Interchange-fee regulation now anchors at least three federal appellate cases. Courts will get them wrong unless they recognize a threshold fact: payment-card networks are multisided platforms.

Interchange fees are the amounts card-issuing banks withhold from a transaction before paying merchant-acquiring banks. On a $100 purchase with a 2% interchange fee, the issuing bank keeps $2 and sends $98 to the acquiring bank, which deducts its own fees and remits the rest to the merchant.

Visa and Mastercard set interchange fees to balance the platform’s two sides—merchants and consumers. Issuing banks use the revenue to cover card-program operating costs, including fraud detection and chargeoffs, and to fund cardholder services such as rewards programs and, for debit cards, the costs of maintaining checking accounts.

The Supreme Court recognized the economic logic in Ohio v. American Express:

Sometimes indirect network effects require two-sided platforms to charge one side much more than the other… The optimal price might require charging the side with more elastic demand a below-cost (or even negative) price. With credit cards, for example, networks often charge cardholders a lower fee than merchants because cardholders are more price sensitive. In fact, the network might well lose money on the cardholder side by offering rewards such as cash back, airline miles, or gift cards. The network can do this because increasing the number of cardholders increases the value of accepting the card to merchants and, thus, increases the number of merchants who accept it. Networks can then charge those merchants a fee for every transaction (typically a percentage of the purchase price). Striking the optimal balance of the prices charged on each side of the platform is essential for two-sided platforms to maximize the value of their services and to compete with their rivals. [citations omitted].

Because payment-card markets are two sided, regulating interchange fees affects far more than banks and merchants. Evidence from the Durbin amendment’s debit-card price controls illustrates the point. After the law capped interchange fees for debit cards issued by larger banks, banks cut debit rewards, sharply reduced free-checking availability, and raised minimum balances needed to avoid maintenance fees. In the wake of those changes, hundreds of thousands of consumers left the banking system. Merchants, meanwhile, passed through little—if any—of the savings as lower retail prices.

Read the full piece here.

Abstract

This article questions the legal validity of the Digital Markets Act (“DMA”) in light of its enforcement practice. Adopted on the basis of Article 114 TFEU as an internal market harmonization measure, the DMA is administered by the Commission as a standing regime of unilateral conduct control that operates alongside, and in close normative proximity to, Article 102 TFEU. The resulting functional equivalence between the two instruments raises structural doubts as to the DMA’s compatibility with the constitutional framework of the Treaties.

View on SSRN.

Executive Summary

The International Center for Law & Economics (ICLE) submits these comments on the Competition Bureau’s Draft Merger Enforcement Guidelines. The Draft Merger Guidelines consolidate the Bureau’s merger-review approach following amendments to the Competition Act enacted through Bill C-56 (2023) and Bill C-59 (2024). Those amendments removed the efficiencies defence, introduced structural presumptions based on concentration and market share, authorised consideration of labour-market effects, and expanded private access to the Competition Tribunal.

These comments identify three principal concerns and propose targeted revisions.

The Draft Guidelines presume a substantial lessening or prevention of competition where the post-merger HHI exceeds 1,800 (with an increase of 100 or more) or the parties’ combined market share exceeds 30%. These thresholds track the 2023 U.S. Merger Guidelines, which were criticised by the competition-enforcement agencies’ own economists as lacking empirical support. Decades of industrial-organization research—including surveys by Steven Berry, Fiona Scott Morton, and Martin Gaynor, and work by former chief economists at both U.S. enforcement agencies—show that concentration measures correlate poorly with competitive harm across industries and market conditions.

Concentration often reflects competitive success. Firms with lower costs or better products gain share as higher-cost rivals lose sales, increasing concentration even as margins decline. The Draft Guidelines also leave unclear how parties may rebut the presumption and how it applies to primarily vertical or conglomerate transactions.

The Bureau should treat structural presumptions as preliminary screens, not near-determinative findings of harm. Below a 30% combined market share, enforcement should be exceptional. Below an HHI of 2,500, presumptions should be readily rebutted. The guidelines should identify rebuttal evidence—entry conditions, rival expansion, switching costs, multi-homing, innovation, merger-specific efficiencies, and direct competitive-effects evidence—and clarify application to non-horizontal transactions.

Following repeal of the former s. 96 efficiencies defence, the Draft Guidelines state efficiencies are “unlikely to change” the Bureau’s conclusions when a merger raises significant concerns. This framing treats efficiencies as peripheral, rather than central to competitive effects. Standard price theory predicts the opposite: marginal-cost reductions directly affect pricing and output. A merger that sufficiently reduces marginal cost may lower post-merger prices even if concentration increases. Dynamic efficiencies—innovation, product quality, and faster commercialisation—often generate substantial consumer benefits in research-intensive sectors.

The issue is particularly important for vertical mergers. The elimination of double marginalisation—the pricing distortion that occurs when firms with pricing power operate at successive levels of a supply chain—is an inherent consequence of vertical integration, not a speculative synergy. Empirical literature consistently finds vertical mergers benefit consumers; a survey of 31 studies found consumer benefit in all but one. Yet the Draft Guidelines apply horizontal structural thresholds to vertical transactions, assert without citation that vertical integration increases coordination risk, and subject double marginalisation elimination to demanding evidentiary requirements. The draft also applies inconsistent reasoning to contractual alternatives, treating contracts as insufficient to prevent foreclosure but sufficient to replicate integration benefits.

The Bureau should integrate efficiencies directly into competitive-effects analysis, apply a realistic merger-specificity standard that accounts for timing and execution costs, and recognise that substantial efficiencies can rebut structural presumptions. For vertical mergers, the Guidelines should establish a safe harbour where shares remain below 40% at each level, recognise elimination of double marginalisation where both levels exercise pricing power, remove the unsupported coordination-risk claim, apply consistent treatment of contracts, and intervene only where foreclosure would harm consumers and impair rivals’ ability to compete.

Several provisions extend enforcement theories into areas where economic evidence is thinner and conventional tools are less reliable. In platform markets, the Draft Guidelines risk treating network effects as a presumption of market power. Network effects vary in strength and scope and must be assessed alongside entry conditions, switching costs, innovation, and multi-homing. Platform markets frequently experience rapid displacement by new entrants.

The Bureau should specify that widespread multi-homing and low switching costs ordinarily weigh against a finding of substantial lessening of competition, require concrete evidence that a target is a likely and substantial future competitive constraint before concluding a merger prevents dynamic competition, and evaluate platform-participant mergers under the vertical-effects framework.

The labour-market provisions raise similar concerns. The Draft Guidelines apply a product-market analytical framework to labour markets, even though labour-market definition is more uncertain and the economic literature remains unsettled. Reductions in labour demand can reflect efficiency gains rather than monopsony power, and the two cannot be distinguished without examining downstream output and prices.

The Bureau should require evidence of downstream output effects in labour-market analysis, apply a demanding evidentiary standard, and define labour markets to reflect realistic worker alternatives, including cross-industry substitution and remote-work opportunities.

An error-cost framework links these recommendations. False positives—blocking mergers that would benefit consumers—impose durable costs: forgone efficiencies, deterred transactions, and reduced investment and innovation. False negatives are more likely to self-correct through entry, expansion, and technological change. The Guidelines should therefore direct enforcement toward mergers supported by substantial evidence of harm and exercise restraint where evidence is ambiguous or offsetting benefits are likely. The proposed revisions would reduce false positives while preserving strong enforcement against mergers that demonstrably harm competition and consumers.

I. Introduction and Overview

The International Center for Law & Economics (ICLE) submits these comments on the Competition Bureau’s Proposed Merger Enforcement Guidelines.[1] ICLE is a non-profit research centre that applies economic analysis to legal and regulatory questions, with a particular focus on competition policy and market institutions. ICLE scholars previously submitted comments to the Bureau on the Future of Competition Policy in Canada,[2] on Canada’s Anti-Competitive Conduct and Agreements Enforcement Guidelines (attached as Appendix AA),[3] and on the Bureau’s prior merger-enforcement guidelines (attached as Appendix BB).[4]

The Draft Merger Guidelines aim to consolidate the Bureau’s merger-review approach following significant amendments to the Competition Act enacted through Bill C-56 (2023) and Bill C-59 (2024).[5] Those amendments materially expanded the Bureau’s enforcement authority. Among other changes, Parliament:

• removed the efficiency defence in merger review (Bill C-56);
• introduced a presumption that a merger is anti-competitive if it significantly increases concentration or market share (Bill C-59);
• strengthened remedies by directing that they preserve or restore the level of competition that would have existed absent the merger (Bill C-59);
• clarified that the Competition Tribunal may consider competitive harm in labour markets when deciding whether to allow a merger or part of a merger (Bill C-59); and
• clarified that the Tribunal may consider coordination risk among competitors (Bill C-59).

Bill C-56 also added “excessive and unfair selling prices” as an anti-competitive act under s. 78, extended s. 90.1’s civil-agreements provisions to non-competitors, and increased maximum administrative monetary penalties to the greater of $10 million for individuals or $25 million for corporations, or up to three times the benefit derived from the conduct. Bill C-59 expanded private access to the Competition Tribunal. ICLE addressed those issues in the ACCA Comments submitted in January 2026 (Appendix AA).

Although Parliament broadened the statutory language, the Bureau retains substantial discretion in setting enforcement priorities and interpretive standards. These comments address how the Bureau can exercise that discretion to protect competition without discouraging mergers that benefit consumers.

A. Structural Presumptions

The Draft Merger Guidelines place substantial weight on structural presumptions that economic research shows are unreliable predictors of competitive harm. The proposed thresholds—a Herfindahl–Hirschman index (HHI) of 1,800 with an increase of 100, or a combined market share exceeding 30%—lack empirical support.

Decades of industrial-organization research, including work by economists who have served in U.S. enforcement agencies, show that concentration measures cannot substitute for analysis of entry conditions, firm conduct, and competitive dynamics. The proposed thresholds closely track the controversial 2023 U.S. Merger Guidelines, which agency economists themselves criticised as unsupported by the empirical literature.

The draft further leaves unclear how merging parties may rebut the presumption and how the presumption applies to primarily vertical or conglomerate transactions.

B. Efficiencies

The Guidelines marginalise efficiencies in a manner inconsistent with standard price theory and likely to generate systematic false positives. After Parliament repealed the s. 96 efficiency defence, the Guidelines treat efficiencies as unlikely to affect the Bureau’s conclusions whenever a merger raises structural concerns.

This approach disregards the role of cost reductions in shaping firms’ pricing and output decisions. The problem is particularly acute for vertical mergers. Eliminating double marginalisation is a direct consequence of integration, not a speculative synergy, and empirical literature consistently finds consumer benefits.

The Guidelines also adopt internally inconsistent assumptions about contracts. They treat contracts as insufficient to prevent foreclosure harms, yet sufficient to replicate the benefits of integration.

C. Expanding Theories of Harm

Several provisions extend enforcement theories into areas where economic evidence is limited or the analytical framework fits poorly. The claim that vertical integration increases coordination risk appears without citation or economic reasoning. The treatment of platform mergers risks turning network effects into a presumption of market power, rather than evaluating them alongside entry, switching costs, and multi-homing.

The labour-market provisions apply a product-market framework to settings where market definition is more difficult and the literature less developed. They also risk misidentifying efficiency-enhancing reductions in labour demand as monopsony power, unless downstream output effects are examined.

D. Organisation of These Comments

These comments proceed as follows. Section II explains the error-cost framework and its implications for the Bureau’s exercise of enforcement discretion. Section III examines the structural presumptions and concentration thresholds, including the limited predictive value of HHI measures, the specific numerical thresholds adopted, and the unclear rebuttal standards. Section IV addresses market definition and its proper analytical role. Section V discusses efficiencies and the failing-firm defence. Section VI analyses vertical mergers, including empirical evidence on competitive effects, the unsupported coordination-risk claim, foreclosure theories and their limits, and the mischaracterisation of double marginalisation as a speculative efficiency. Section VII examines platforms and multi-sided markets, including multi-homing, dynamic competition, and self-preferencing. Section VIII addresses labour-market analysis and the gap between the economic literature and the proposed framework. Section IX sets out specific recommendations for revision.

II. The Error-Cost Framework and Effects-Based Merger Analysis

As noted in ICLE’s ACCA Comments, competition enforcement necessarily operates under uncertainty and therefore carries a risk of error. The error-cost framework developed by U.S. Judge Frank Easterbrook and subsequent scholarship recognises that enforcement rules must account for both over- and under-deterrence.

Under that framework, the social costs of false positives (Type I errors) generally exceed those of false negatives (Type II errors). Erroneous intervention can chill pro-competitive conduct, deter innovation, and create lasting legal and economic distortions that market forces cannot readily correct. By contrast, under-enforcement is more often temporary, as entry and competitive pressure tend to erode supra-competitive outcomes over time.

These considerations favour clear limiting principles, rigorous evidence, and interpretive restraint in the Bureau’s enforcement approach, particularly in dynamic or innovation-driven markets, where over-enforcement produces the greatest harm.

III. Structural Presumptions and Concentration Thresholds

The Guidelines adopt the structural presumptions in s. 92(2) of the Competition Act. A merger that significantly increases market share or concentration is presumed to substantially lessen or prevent competition unless the parties rebut the presumption. Under s. 92(3), a “significant increase” occurs when:

• the post-merger HHI rises, or is likely to rise, by more than 100; and
• either the concentration index exceeds, or is likely to exceed, 1,800, or the parties’ combined post-merger market share exceeds, or is likely to exceed, 30%.

The Guidelines state that mergers meeting these thresholds are “presumed to substantially harm competition.” Section 92(2) allows the parties to rebut the presumption on a balance of probabilities.

Market-share measures can appropriately serve as an initial screening tool in merger review. The proposed approach, however, raises several concerns.

First, decades of economic research show that market shares and concentration ratios do not reliably predict competitive harm. The structural presumptions therefore place excessive weight on concentration measures.

Second, the numerical thresholds are too low and lack a sound empirical basis across diverse markets, particularly where the Guidelines attach a presumption of competitive harm.

Third, the Guidelines do not clearly specify the evidentiary standard required to rebut the presumption.

Fourth, the Guidelines do not explain how the presumptions apply to primarily vertical or conglomerate transactions that include limited horizontal overlap.

A. The Limited Predictive Value of Concentration Measures

Concentration measures, and market-share thresholds in particular, have limited value in predicting competitive effects. They may serve as preliminary screening tools and, in some markets, they may signal competitive concerns. Decades of economic research nonetheless show that market shares and concentration ratios do not reliably predict competitive harm across industries or sectors. A market with many small firms is not necessarily more competitive than one with fewer large firms. The proposed structural presumptions therefore place excessive weight on concentration measures.

Indeed, concentration often reflects competitive success, rather than market failure.[6] Firms with lower costs or better products expand, while higher-cost rivals lose sales.[7] As Chad Syverson observes, “many empirical studies in varied settings have found that greater substitutability/competition—resulting from, say, reductions in trade, transport, or search costs—shifts activity away from smaller, higher-cost producers and toward larger, lower-cost producers,” increasing concentration even as margins decline.[8]

Treating “too much” concentration as inherently harmful assumes both that market structure determines economic outcomes and that a correct level of concentration can be identified. Economists have rejected that premise since at least the 1970s. As one classic formulation explains:[9]

Once perfect knowledge of technology and price is abandoned, [competitive intensity] may increase, decrease, or remain unchanged as the number of firms in the market is increased.… [I]t is presumptuous to conclude… that markets populated by fewer firms perform less well or offer competition that is less intense.[10]

Harold Demsetz’s critique of the “market concentration doctrine” likewise showed that correlations between concentration and profits often reflect efficiency, rather than market power.[11] A recent survey of the industrial-organization literature by Steven Berry, Fiona Scott Morton, and Martin Gaynor reached a similar conclusion:

In short, there is no well-defined “causal effect of concentration on price,” but rather a set of hypotheses that can explain observed correlations of the joint outcomes of price, measured markups, market share . . .

Our own view, based on the well-established mainstream wisdom in the field of industrial organization for several decades, is that regressions of market outcomes on measures of industry structure like the Herfindahl-Hirschman Index should be given little weight in policy debates.^[12]

Concentration measures therefore provide, at most, rough starting points for analysis. Dennis Carlton describes market shares and concentration as “at best crude first steps,” noting a weak empirical link between changes in market share and competitive performance, and observing that no model shows market share alone reliably predicts prices or welfare effects.[13]

Experience in health-care merger enforcement illustrates the point. U.S. enforcement agencies have relied on applied industrial-organization research to challenge anticompetitive hospital mergers successfully in court, while moving beyond traditional structure-conduct-performance (SCP) assumptions. Agency research finds that newer screening tools—particularly willingness-to-pay (WTP) and upward pricing pressure (UPP)—more accurately identify problematic hospital mergers than traditional concentration measures.[14]

Static market-share data cannot substitute for analysis of firm conduct and market conditions. Modern competition analysis instead emphasises entry conditions and potential competition. The contestable-markets framework developed by William Baumol, John Panzar, and Robert Willig shows that industry outcomes depend primarily on entry and exit costs, particularly sunk costs, rather than on the number or size of incumbent firms. Even concentrated markets can remain competitive where entry is feasible, because “potential competition enforces cost-minimization regardless of the nature of the industry structure.”[15] David Teece extends this insight to innovation-driven sectors, arguing that market share poorly captures market power when technological change and “unseen” competition constrain incumbents.[16] OECD methodological reviews likewise emphasise contestability, entry barriers, and switching costs—factors not reflected in standard concentration metrics.[17]

Evidence from productivity[18] and innovation[19] studies reinforces this conclusion. Competitive pressure tied to entry and policy changes correlates more strongly with performance than market structure alone. Authorities should therefore treat market shares as incomplete and potentially misleading screens, supplemented—and, where appropriate, displaced—by analysis of entry, expansion, switching costs, innovation, and the durability of any competitive advantage.

Digital markets illustrate these dynamics. Firms such as Facebook, Google, and Amazon achieved large market shares by offering valuable services at low or zero prices and improving user experience. Their positions remain contestable through innovation, as reflected in TikTok’s rapid growth in social media and continued competition in e-commerce from Shopify, Walmart, and specialised retailers.

The final Guidelines should clarify that market-share thresholds function only as initial screens, not presumptions of market power. Below 30%, the Bureau should decline enforcement absent exceptional circumstances. Above 30%, the Guidelines should identify factors—such as entry conditions, rival expansion, customer switching, and innovation—that can rebut any inference of market power drawn solely from market-share data.

B. Structural Presumptions Require Clear Rebuttal Standards

Section 92(2) creates a structural presumption that shifts the burden to the merging parties to show the transaction is unlikely to substantially lessen or prevent competition. The Guidelines add that “the more the thresholds are exceeded, the greater the need for persuasive evidence to refute the presumption.”

As explained in ICLE’s ACCA Comments, the Bureau should clarify that structural presumptions function only as initial screens, not prima facie—and potentially determinative—evidence of competitive harm to Canadian consumers. Market definition, market shares, and concentration measures are analytical tools, not conclusions. The Bureau should evaluate them alongside:

• entry conditions and the likelihood of timely, likely, and sufficient entry;
• expansion by existing rivals;
• customer switching costs and multi-homing behaviour;
• innovation dynamics and the pace of technological change; and
• direct competitive-effects evidence, including customer testimony, internal documents, and economic analysis.

The proposed Guidelines also do not specify the evidentiary standard required to rebut a structural presumption. That omission is particularly concerning, given the limited empirical support for the presumptions themselves. Without clear standards, parties cannot predict what evidence will suffice, and enforcement risks becoming discretionary rather than evidence-based.

C. The Proposed HHI Thresholds Lack Empirical Support

The Guidelines adopt an HHI threshold of 1,800, with an increase of 100 or more, to trigger the structural presumption. This threshold appears drawn from the U.S. agencies’ 2023 Merger Guidelines,[20] which departed from prior enforcement practice.

The U.S. agencies’ 2010 Horizontal Merger Guidelines identified three concentration ranges: “unconcentrated markets” (HHI below 1,500), “moderately concentrated markets” (HHI between 1,500 and 2,500), and “highly concentrated markets” (HHI above 2,500). Mergers producing a moderately concentrated market with an HHI increase above 100 were deemed to “potentially raise significant competitive concerns” and “often warrant scrutiny.”[21] Even mergers producing a highly concentrated market were not presumed unlawful. Instead, mergers that increased HHI by 100–200 points were likewise said to “potentially raise significant competitive concerns and often warrant scrutiny.”[22]

The 2023 U.S. revisions proved controversial and departed from both historical agency practice and mainstream economic literature.[23] Agency economists themselves recognised longstanding concerns about overreliance on concentration metrics such as HHIs. In 2022, Steven Berry, Fiona Scott Morton, and Martin Gaynor joined numerous former chief economists from both U.S. competition agencies, as well as Aviv Nevo (soon to become director of the FTC’s Bureau of Economics), in cautioning that “regressions of price on the HHI do not predict the competitive effects of mergers and should not be used in merger review.”[24]

The point was not that increasing concentration can never justify scrutiny or that HHI lacks screening value. Rather, relative HHI changes do not reliably predict competitive harm—or benefit—and therefore should not be embedded in legal presumptions.

The specific numerical revisions adopted in 2023 also lacked support. In comments submitted to the U.S. agencies before adoption of the guidelines, Aviv Nevo and other prominent economists warned:

Our view is that this would not be the most productive route for the agencies to pursue to successfully prevent harmful mergers, and could backfire by putting even further emphasis on market definition and structural presumptions.

If the agencies were to substantially change the presumption thresholds, they would also need to persuade courts that the new thresholds were at the right level. Is the evidence there to do so? . . . Our reading of the literature is that it is not clear and persuasive enough . . . to support a substantially different threshold that will be applied across the board to all industries and market conditions.[25]

ICLE scholars and others likewise criticised the U.S. revisions in comments to the U.S. Justice Department (DOJ) and Federal Trade Commission (FTC).[26] The lower thresholds lack empirical support and risk capturing many mergers that pose no competitive threat. As Nevo, John Asker, and other co-authors observed, the evidence supporting structuralist revisions was “thin.”[27] Economic evidence provides no reliable basis to treat an HHI of 1,800 or a 30% market share as a dependable indicator of market power or likely competitive harm.

The Competition Bureau should therefore avoid importing and extending these unsupported presumptions. Under an error-cost framework, the relevant choice is not between more and less aggressive enforcement, but between evidence-based and presumption-based enforcement. Poorly justified presumptions risk increased false positives without offsetting benefits, harming competition in Canada and Canadian consumers.

The Guidelines should clarify three points.

First, HHI calculations depend on market definition, which is itself an analytical construct subject to uncertainty and judgement. Where market boundaries are contested or unclear, structural presumptions based on HHI should carry less weight.

Second, HHI measures static market structure. It does not capture dynamic competition, innovation, potential entry, or competitive self-correction. High HHIs in rapidly changing markets may reflect temporary leadership rather than durable market power.

Third, below an HHI of 2,500, structural presumptions should apply cautiously and should be readily rebutted by evidence of entry, expansion, innovation, or efficiencies.

IV. Market Definition and Competitive-Effects Analysis

Market definition is a central analytical tool in merger review, but it presents significant conceptual and practical challenges. This section evaluates the Guidelines’ approach to defining relevant product and geographic markets and treats market definition as a flexible method for assessing competitive effects, not a rigid requirement.

The discussion reviews the strengths and limits of the hypothetical monopolist (SSNIP) test, including the Cellophane fallacy, critical-loss circularity, non-price competition, and product differentiation. It also addresses the factors that shape geographic market boundaries. The section concludes with recommendations to better align the Guidelines with economic principles and the practical realities of competitive analysis.

A. Market Definition as an Analytical Tool

The Guidelines correctly recognise that market definition is an analytical tool, not an end in itself. As the Guidelines state, “in some reviews, we may not precisely define markets, because we are able to assess the merger’s effects on competition without doing so.”

This flexibility reflects sound economic principles. Market definition draws a boundary between closer and more distant substitutes. In many cases, precise boundaries are unnecessary and sometimes unattainable. Buyers’ close substitutes will not always fall neatly inside a single market, and competitive constraints may arise from both inside and outside any defined boundary.

The Guidelines should emphasise this point more clearly. Market definition serves one purpose: assessing competitive effects. Where the Bureau can evaluate competitive effects directly, market definition becomes less important. Where merging firms compete across multiple overlapping markets with similar competitive conditions, aggregating those markets may simplify analysis without affecting the result.

B. The Hypothetical Monopolist (SSNIP) Test and Its Limits

The Guidelines describe the hypothetical monopolist test—also known as the SSNIP test (a “small but significant and non-transitory increase in price”)—as a tool for defining markets. The test asks whether a hypothetical monopolist controlling all products in a candidate market would profitably impose a price increase. The Guidelines specify a 5% increase lasting one year.

The SSNIP framework offers useful conceptual guidance but presents well-known implementation challenges.

1. The Cellophane Fallacy

If current prices already reflect market power, using them as the SSNIP baseline will overstate the relevant market. A monopolist already charging supra-competitive prices would lose substantial sales from an additional 5% increase, incorrectly suggesting a broader market. The Guidelines recognise this concern, noting that “we may use a different base price when market conditions absent the merger would likely result in a lower or higher price in the future.”

2. Critical-loss analysis

Applying the SSNIP test often requires estimating “critical loss”—the maximum quantity loss a hypothetical monopolist could sustain while still finding a price increase profitable. Critical loss depends on margins, which themselves depend on market definition. This creates circularity. High observed margins may indicate market power, implying narrow markets, or efficiency, which does not.

3. Non-price competition

The SSNIP test focuses on price, yet many mergers affect quality, variety, innovation, privacy, or other non-price dimensions. The Guidelines acknowledge this, stating that “depending on the nature of competition, we may assess whether a different exercise of market power would be profit-maximizing, such as a change in product quality, variety, privacy, or any other dimension of competition.” The final Guidelines should provide more specific guidance on applying the hypothetical monopolist framework to non-price competition.

4. Product differentiation

In differentiated-product markets, the SSNIP test may produce very narrow markets, potentially as narrow as a single product. Competitive effects instead depend on substitution patterns and the degree of differentiation. The Guidelines should clarify that diversion ratios, customer-switching evidence, and other direct competitive-effects evidence often provide more reliable guidance than formal market definition in such settings.

C. Defining Geographic Markets Based on Competitive Constraints

The Guidelines apply the hypothetical monopolist test to geographic markets. If buyers would shift purchases to sellers in more distant locations in quantities sufficient to make a price increase unprofitable, those locations become part of the candidate market. The approach is sound in principle but often difficult to implement.

Several factors are particularly important.

• Transportation Costs: High transportation costs limit the geographic scope of competition. Bulk commodities and heavy products typically compete in regional, not national, markets.
• Regulatory Barriers: Provincial or national regulations may restrict out-of-market suppliers. Licensing requirements, product-approval processes, and language requirements can define practical geographic boundaries.
• Customer Behaviour: Actual purchasing patterns, customer location, and willingness to travel or source from distant suppliers provide direct evidence of geographic market scope.

The Guidelines should emphasise that geographic markets must reflect real competitive constraints, not administrative or political borders. Where customers regularly source from multiple provinces or countries, those areas should fall within the geographic market even if suppliers maintain separate regional pricing policies.

V. Countervailing Considerations: Efficiencies and Failing Firms

This section addresses two countervailing considerations identified in the Guidelines: efficiencies (Part 4.7) and failing firms and exiting assets (Part 4.6). Both bear directly on whether a merger causes a substantial lessening or prevention of competition and both warrant clearer, more economically grounded treatment in the final Guidelines.

First, although Parliament repealed the former s. 96 efficiency defence, efficiencies remain central to competitive-effects analysis. Cost reductions, innovation gains, and other merger-specific efficiencies shape firms’ pricing, output, and investment decisions and may offset, or more than offset, any increase in concentration. The Guidelines should therefore integrate substantiated efficiencies into the SLOC analysis, rather than treat them as peripheral or unlikely to affect the Bureau’s conclusions.

Second, the failing-firm and exiting-assets framework properly focuses on causation. Where, absent the merger, assets would leave the relevant market or cease to impose a meaningful competitive constraint, the merger does not worsen competitive outcomes. The analysis should therefore emphasise realistic counterfactuals, commercially reasonable searches for alternatives, and the likelihood of asset exit, including the failure of a part of a business, rather than formal insolvency alone.

Taken together, these considerations ensure merger review evaluates actual competitive effects. The Bureau should intervene where a transaction causes a material loss of competition, not where efficiencies benefit consumers or where assets would otherwise exit the market.

A. Efficiencies as a Component of Competitive-Effects Analysis

The Draft Merger Guidelines reflect the Bureau’s interpretation of Parliament’s repeal of the efficiency defence, formerly codified in s. 96 of the Competition Act. That repeal marks a significant shift in Canadian merger policy. Under the prior provision, a merger that increased prices could still proceed if efficiency gains outweighed the associated deadweight loss. The statutory balancing test has disappeared, but the economic logic underlying it has not.

In Canada (Commissioner of Competition) v. Superior Propane Inc.,^[28] the Federal Court of Appeal explained that a merger producing a substantial lessening of competition could be permitted where efficiency gains exceeded and offset the anti-competitive effects. The Court discussed the “total surplus standard,” which focuses on deadweight loss, while also directing the Tribunal to consider the Act’s broader purposes, including competitive prices for consumers.

Under the amended Act, efficiencies appear only as one factor in the non-exhaustive list relevant to a substantial lessening or prevention of competition (SLOC).^[29] The Draft Guidelines state the Bureau will consider efficiencies “as part of our overall analysis of harm to competition only where they are clearly merger-specific, substantiated with rigorous and independent evidence, and demonstrably likely to enhance competitive outcomes in a way that benefits Canadians.”^[30] The Guidelines also state efficiencies are “unlikely to change” the Bureau’s conclusion where a merger raises significant competition concerns.^[31]

Taken literally, this approach risks treating efficiencies as peripheral, rather than central, to competitive effects. Standard price theory places cost conditions at the core of pricing incentives, output decisions, and responses to entry.^[32] As ICLE explained in its earlier comments:

And yet it, is not entirely clear—in the aftermath of Bill C-59, and especially following the elimination of the efficiencies defense—how defendants could rebut the structural presumption laid down in s.92, other than by refuting the market-concentration calculus put forward by the Bureau. This essentially turns merger review under the Competition Act into a formalistic exercise where—despite assurances to the contrary—market structure is outcome-determinative. It also jars with the logic that applies to conduct under s. 90.1, thereby rendering the Competition Act conceptually rudderless.[33]

1. The economic role of efficiencies

Discounting efficiencies risks systematic error. Enforcement that prioritises structural indicators while downplaying cost reductions invites false positives by condemning transactions that may reduce prices or expand output. The Draft Guidelines move policy in that direction by suggesting efficiencies rarely affect the SLOC determination once market-power concerns arise.

The relevant question is not whether efficiencies “preserve or restore competition to a level that would exist without the merger,” but whether they improve consumer welfare despite any reduction in rivalry. A merger may eliminate a competitor yet still benefit consumers if cost savings outweigh any price effects.

Efficiencies also matter at the macroeconomic level. Mergers that eliminate duplicative fixed costs, rationalise production, or achieve economies of scale or scope free resources for use elsewhere in the economy. For a small, open economy such as Canada’s, scale efficiencies may enable domestic firms to compete against larger foreign rivals.

The Guidelines should integrate efficiencies directly into competitive-effects analysis, rather than treat them as a secondary step after identifying an SLOC. Marginal-cost reductions are central to unilateral-effects analysis. If a merger lowers marginal costs sufficiently, the profit-maximising post-merger price may fall relative to pre-merger prices, even when concentration increases.

Joseph Farrell and Carl Shapiro show that whether efficiencies lower prices will depend on the merging firms’ market shares and demand elasticity.^[34] A merger that lowers marginal costs will often lead the merged firm to reduce prices and expand output. The Guidelines should therefore evaluate efficiency claims carefully, not discount them based on structural indicators alone.

2. Dynamic and fixed-cost efficiencies

Dynamic efficiencies warrant similar attention. Improvements in innovation, product quality, and speed of commercialisation often generate large consumer benefits, especially in research-intensive industries. Merger review focused narrowly on short-run price effects risks overstating competitive harm and understating these gains. A merger that accelerates innovation or combines complementary assets can intensify competition over time, even if concentration increases in the short run. Gregory Sidak and David Teece argue that a “neo-Schumpeterian” approach places less weight on market share and more on potential competition and firm capabilities.^[35]

Fixed-cost savings may not immediately reduce prices, but they affect investment incentives, risk tolerance, and capacity expansion. Ignoring fixed-cost efficiencies biases analysis toward static price effects and away from long-run competition. Oliver Williamson demonstrated that merger analysis should weigh scale economies against deadweight loss to avoid significant economic waste.^[36]

3. Buyer-power efficiencies

The categorical exclusion of cost savings from stronger bargaining with suppliers lacks economic support. Lower input prices obtained through negotiation constitute an efficiency unless they reflect monopsony power that reduces output below competitive levels. The OECD notes that monopsony depresses output and prices below competitive levels, while bargaining power can countervail seller market power and move prices toward competitive levels.^[37]

Economic theory distinguishes between redistributive transfers and true cost reductions that affect downstream pricing incentives.^[38] Before Bill C-56, the Bureau recognised this distinction in reporting to the OECD, noting that it assessed whether claimed efficiencies merely redistributed income between parties.^[39]

4. Merger specificity

The Guidelines require efficiencies to be “merger-specific.” The economic literature shows that the choice between building and buying capabilities is not costless.[40] Peter Klein and João Mazzoni find firms vertically integrate to protect relationship-specific investments and resolve conflicts arising from incomplete contracts. Integration may therefore create efficiencies—such as protecting specific assets—that arm’s-length contracts cannot replicate.

Organic expansion takes time, requires capital, and involves execution risk. A merger may enable faster integration, avoid duplication, and transfer tacit knowledge that internal development cannot easily reproduce.

Assuming contracts can replicate integration efficiencies lacks theoretical[41] and empirical support.[42] If a firm could easily replicate a target’s capabilities internally, it would not pay an acquisition premium. The transaction itself provides market-based evidence that the merger creates value not achievable at comparable cost or speed. The Guidelines should compare realistic alternatives, not a hypothetical first-best scenario that ignores timing, risk, and implementation costs.

5. Interaction with structural presumptions

The Guidelines do not explain how efficiencies interact with the structural presumption under s. 92(2). If a merger exceeds concentration thresholds, can substantial efficiencies rebut the presumption, or must parties still prove on a balance of probabilities that no SLOC will occur?

The Guidelines should clarify:

• Efficiencies are directly relevant to whether a merger substantially lessens or prevents competition.
• Parties may rebut the structural presumption by demonstrating that efficiencies offset potential harm.
• Where efficiencies are large and merger-specific, the Bureau should not require unattainable precision in quantification. Merger analysis is inherently predictive, and the same uncertainty affecting competitive-effects analysis affects efficiency measurement.

The repeal of s. 96 changed the legal framework, not the economics of competition. The Guidelines should treat substantiated efficiencies as integral to the SLOC analysis under s. 93, rather than as a rare exception to structural concerns. Efficiencies benefit consumers, not “competition” in the abstract. A merger that eliminates duplication, integrates complementary assets, or achieves scale can benefit consumers even if it reduces the number of firms in a narrowly defined market.

B. Failing Firms, Exiting Assets, and the Counterfactual

Section 4.6 of the Draft Merger Guidelines addresses failing firms and exiting assets. It reflects a basic economic principle: if, absent the merger, a firm’s assets would leave the relevant market and cease to impose a material competitive constraint, the merger does not cause any incremental loss of competition. The draft recognises this point, stating that parties may show “the merger is not the cause of the loss of a competitor” where, without the transaction, “imminent failure is probable and the assets of the firm are likely to exit the relevant market and no longer materially impact competition in that market.”[43] This causation focus accords with s. 93(b) of the Competition Act, which asks whether “the business, or a part of the business, of a party to the merger or proposed merger has failed or is likely to fail.”

The Draft Guidelines emphasise that “probable business failure does not provide a defence for a merger that is likely to prevent or lessen competition substantially.”[44] Instead, financial distress and asset exit function as rebuttal evidence on causation. A categorical defence could immunise harmful mergers where less harmful alternatives exist. At the same time, if the counterfactual is asset exit and loss of competitive constraint, blocking the merger would not improve competitive outcomes.

1. Insolvency requirements versus asset exit

The draft defines a “failing firm” narrowly: a firm that is insolvent or likely to become insolvent, has initiated or is likely to initiate bankruptcy proceedings, or faces receivership.[45] It also requires proof the firm cannot reorganise successfully under insolvency regimes or voluntary arrangements.[46]

This approach may be too restrictive. Section 93(b) refers to “the business, or a part of the business.” Many transactions involve business units, facilities, or product lines that would exit the relevant market even if the corporate parent remains solvent.

Economically, the relevant question is whether the assets that constrain competition would continue to do so absent the merger. A plant, network, or product line likely to shut down or redeploy outside the relevant market no longer disciplines rivals, even if the firm as a whole remains financially viable.[47] The Guidelines should clarify that failure or exit of a “part of the business”—such as a division, line of business, or local operation—can satisfy this element, provided the parties show those assets would cease materially affecting competition.

2. Alternative purchasers and commercial reality

The Draft Guidelines require “reliable and objective evidence” of financial distress and state that “subjective expectations or unsupported projections will not be considered credible.”[48] That standard is appropriate. Parties hold superior information and may have strategic incentives.

Predicting exit nonetheless involves uncertainty. The Guidelines should not require certainty. A workable standard is whether contemporaneous financial information, internal documents, and third-party evidence show exit is more likely than continued operation at the current competitive level absent the merger.

Section 4.6 also requires proof that “no alternative to the merger [is] likely to result in a materially greater level of competition.”[49] It identifies three alternatives: acquisition by a competitively preferable purchaser (CPP), retrenchment or restructuring, and liquidation.[50] The draft asks whether another purchaser could produce a materially higher level of competition and pay more than liquidation value plus transaction costs. It also contemplates requiring an independent third party to search for buyers.[51]

Focusing on counterfactual competitive effects is appropriate. The CPP requirement, though, risks favouring smaller but less efficient firms. As Lars Persson explains, interpreting “competitively preferable” to mean a lower market-share buyer would advantage smaller bidders even when larger firms would generate greater producer and consumer surplus.[52] Smaller firms may also use the condition to block acquisitions that would otherwise increase total surplus.

A stringent “no alternative” requirement combined with mandatory independent searches could impose delay and cost on financially distressed firms. Distressed sellers often retain advisors, canvass strategic and financial buyers, and explore restructuring. The Bureau should evaluate whether the parties conducted a commercially reasonable search given their financial condition, timing constraints, market conditions, and the business’s scale. Presuming unspecified additional searches would uncover a superior buyer risks converting an economic inquiry into a procedural hurdle.

3. Retrenchment, restructuring, and liquidation

The Draft Guidelines recognise that retrenchment—downsizing or withdrawing from certain products or regions—may allow a firm to survive as a competitor.[53] The key question is whether such a strategy would preserve materially greater rivalry in the relevant market than the proposed merger. The Guidelines should focus on whether retrenchment maintains effective competitive constraints, not whether it preserves the firm as a legal entity or employer.

The discussion of liquidation notes that it may “facilitate entry into a market” by allowing competitors to acquire customers or assets.[54] The effect depends on asset characteristics and market structure. Where assets are fungible and easily redeployed—such as standardised equipment—liquidation may support competition. Where assets are highly specific, capable buyers are few, or local demand is weak, liquidation may remove them from the market entirely. The Guidelines should recognise liquidation as superior only when redeployment is reasonably likely to preserve or enhance competition.

4. Clarifying the framework

The Guidelines would be clearer if they explicitly stated that s. 4.6 applies to both firm-wide failure and the failure or exit of a “part of the business” within the meaning of s. 93(b). That clarification would align the Guidelines with the statute and the economic focus on whether specific assets continue to constrain rivals.

Coupled with a likelihood-based evidentiary standard and a commercially realistic assessment of alternatives, this approach preserves the central insight of the failing-firm and exiting-assets analysis: the Bureau should intervene when a merger causes a meaningful loss of competition, not when it merely changes the owner of assets that would otherwise leave the market.

VI. Vertical Mergers and Competitive Effects

A vertical merger joins firms at different levels of a supply chain—for example, a manufacturer acquiring a distributor or a content producer acquiring a distribution platform. The Guidelines correctly centre the analysis on “ability and incentive” to harm competition,[55] and paragraph 230 properly recognises that, where rivals retain “access to many effective alternatives,” substantial harm may be unlikely.[56] Those are appropriate analytical starting points.

The draft nevertheless departs from established economic learning in several important respects. It overstates coordination risks from vertical integration, relies on horizontal-style structural presumptions ill-suited to vertical relationships, and treats efficiencies—particularly predictable benefits of integration—as unlikely to matter in practice. It also frames foreclosure concerns without sufficient attention to realistic counterfactuals and the conditions required for competitive harm.

A large empirical literature finds vertical mergers are commonly pro-competitive, increasing output and benefiting consumers. By setting thresholds and evidentiary standards that discount these effects, the Guidelines risk condemning transactions that improve market performance. The final Guidelines should align vertical-merger analysis with evidence-based competitive-effects principles.

A. Empirical Evidence on Vertical Mergers

A substantial body of economic research finds that vertical mergers typically improve efficiency and benefit consumers. Francine Lafontaine and Margaret Slade surveyed the theoretical and empirical literature on vertical restraints and vertical integration in the Handbook of Industrial Organization and concluded:[57]

[W]e are therefore somewhat skeptical about claims that vertical restraints or vertical integration are likely to harm consumers. In our view, restraints that are imposed should, absent evidence of foreclosure, be assessed under a rule of reason, and firms should almost always be allowed to integrate vertically when that is their choice.[58]

Empirical evidence points in the same direction. The Global Antitrust Institute’s 2018 submission to the U.S. Federal Trade Commission reviewed 31 studies of vertical mergers across multiple industries and found consumer benefits in nearly every case, with “only one study that found evidence vertical integration harmed consumers.”[59] More recent work reinforces this pattern. A 2025 field experiment by Chiara Farronato, Andrey Fradkin, and Alexander MacKay examined Amazon’s private-label products, a form of vertical integration into retail. Using a browser extension that hid Amazon’s own brands from shoppers, the authors found that removing private labels would reduce consumer surplus by 5.4%, with roughly 10% of that reduction attributable to higher prices charged by third-party sellers.[60]

These results reflect basic economic logic. Unlike horizontal mergers among direct competitors, vertical mergers combine firms at different levels of the supply chain and commonly increase output. Manufacturers and distributors both profit from higher sales volumes. Vertical integration often eliminates double marginalisation, reduces free-riding on dealer services, and improves quality assurance.

Benjamin Klein and Kevin Murphy analysed these mechanisms in detail.[61] They show that vertical restraints frequently arise from manufacturers’ efforts to maximise product value to final consumers. Restrictions on distributor conduct can align distributor incentives with consumer welfare, rather than undermine it.

B. The Coordinated-Effects Claim Lacks Economic Support

The Guidelines state that “generally speaking, the greater the degree of vertical integration in a market, the greater the risk of coordinated behaviour.”[62] The draft offers no citation or economic explanation for this proposition, and economic theory does not support it.

Coordinated effects require competitors to reach agreement on coordination terms, monitor compliance, and punish deviation. Vertical integration does not facilitate these mechanisms. Daniel Reiffen and Michael Vita show input foreclosure becomes profitable only when the integrated firm possesses market power at both vertical levels and foreclosure raises rivals’ costs enough to offset lost input sales.[63] Michael Salinger likewise demonstrates foreclosure requires substantial market power both upstream and downstream.[64] Neither condition implies that vertical integration promotes horizontal coordination.

The classic economic insight remains straightforward: vertically related firms have complementary, not collusive, incentives. A manufacturer benefits when distributors sell more units, and a distributor benefits when suppliers produce efficiently. Vertical integration tends to align these incentives, rather than support coordinated output restriction.

The Guidelines should remove the unsupported claim in paragraph 245 or provide economic reasoning and empirical evidence to justify it.

C. Foreclosure Theories in Vertical Mergers

The Guidelines identify foreclosure as the primary concern in vertical mergers. They state a vertical merger may harm competition if it gives the merged firm the ability and incentive to foreclose rivals’ access to an important input or customer base.

This concern is theoretically plausible but often overstated. Foreclosure harms competition only if several conditions hold simultaneously:

• Control of a Critical Input or Customer Base: The acquired business must control access that rivals cannot readily replicate or replace.
• Ability to Foreclose: The merged firm must be able to deny or degrade access. Contracts, regulation, or the loss of profitable upstream sales may make foreclosure infeasible.
• Incentive to Foreclose: Foreclosure must be profitable. If the upstream division earns substantial revenue from selling to rivals, refusing to deal reduces profits. The merged firm will foreclose only if downstream gains exceed lost upstream profits—a condition economic models find uncommon.
• Competitive Harm: Even where foreclosure occurs, consumer harm requires a substantial reduction in downstream competition. If rivals can expand or entry is feasible, foreclosure may disadvantage competitors without harming competition or consumers.

Empirical evidence supports this framework. Studies of vertical integration in cable television, telecommunications, and other sectors generally find neutral or positive effects on prices and output. The theoretical possibility of foreclosure does not justify a presumption against vertical mergers.

D. Structural Presumptions Apply Poorly to Vertical Mergers

The Guidelines apply the same structural presumptions—30% market share and an HHI of 1,800—to vertical mergers as to horizontal mergers. They state the Bureau is “more likely to investigate further” where the merged firm’s share exceeds 30% at either level of the supply chain.[65]

The economics differ fundamentally. A 30% share at one level of a supply chain does not raise the same concerns as a 30% share among direct competitors. Where rivals can obtain inputs or reach customers through alternative trading partners, even a substantial share at one level may pose little competitive risk.

The Guidelines should recognise this distinction. A safe harbour would improve predictability and align enforcement with economic evidence. Vertical mergers should be presumed unlikely to harm competition where the merged firm holds less than a 40% share at each vertical level and rivals have practical access to alternative suppliers or customers.

Below those thresholds, enforcement should require specific evidence of foreclosure that would competitively impair—not merely disadvantage—rivals.

E. Inherent Elimination of Double Marginalisation

The Guidelines treat elimination of double marginalisation as an “efficiency” subject to sceptical evidentiary standards. Footnote 128 states the Bureau “will consider these claims in light of the criteria in paragraph 285, including considering whether any potential elimination of double marginalization is merger-specific, or could be partially or entirely achieved through other contracts, agreements or supply relationships.”[66] This framing misunderstands the economic nature of double marginalisation.

Double marginalisation is a pricing distortion, not a cost. When two firms with pricing power operate at successive levels of a supply chain, each applies a markup. The combined markup exceeds what a single integrated firm would charge. Vertical integration eliminates the distortion by aligning pricing incentives. The upstream division no longer marks up sales to the downstream division because both operate under a unified profit objective.

This distinction has three implications.

• Automaticity: Unlike speculative synergies that depend on post-merger integration, eliminating double marginalisation occurs mechanically upon integration. It requires no reorganisation or managerial effort. Where both levels exercised pricing power pre-merger, the distortion disappears.
• Evidentiary Burden: Applying paragraph 285’s efficiency criteria reverses the proper inquiry. The relevant question is whether both levels possessed pricing power before the merger. If so, elimination follows. The Bureau should explain why the effect would not occur, rather than require parties to prove that it will.
• Contract Alternatives: Footnote 128 suggests contracts or supply agreements could replicate the effect. [67] Yet footnote 106 states “contractual terms are unlikely to eliminate all potential means of foreclosure.”[68] The logic must be consistent. If contracts cannot reliably prevent foreclosure harms, they also cannot reliably reproduce integration benefits. If contracts could perfectly replicate integration, firms would not pay acquisition premiums to integrate.

F. An Unduly High Bar for Pro-Competitive Benefits

Apart from double marginalisation, the Guidelines set an unduly high bar for crediting genuine efficiencies, such as cost savings from synergies, scale, or rationalisation. Paragraph 288 states that “generally, when a merger otherwise presents significant competition concerns, even gains that are supported by rigorous evidence are unlikely to change our conclusions regarding harm to competition.”[69] In practice, this approach risks nullifying efficiency analysis whenever a transaction triggers the structural presumption.

Parliament preserved an efficiencies pathway. Although Bill C-56 repealed the explicit s. 96 defence, s. 93(h) continues to direct the Tribunal to consider “any other factor that is relevant to competition in a market.”[70] Merger-specific efficiencies that enhance consumer welfare are plainly relevant to competitive effects. A framework that recognises efficiencies in principle but discounts them in practice undermines that statutory direction.

VII. Merger Analysis in Multi-Sided Platform Markets

The Draft Merger Guidelines recognise that multi-sided platforms often feature network effects, economies of scale and scope, differentiated monetisation strategies, and dynamic forms of competition.[71] These features make careful analysis especially important. As ICLE noted in its ACCA Comments, however, they also make it critical to anchor enforcement in clear evidence of substantial competitive harm, rather than in broad structural concerns about size, concentration, or business model.

The draft emphasises that network effects and scale economies “may reinforce the growth and market position of incumbents” and in some cases may lead to “tipping.”[72] Network effects are real, but their competitive significance is context-dependent. Users typically value participation by relevant trading partners or contacts, not total global scale, and even strongly networked markets have experienced rapid displacement by entrants. For similar reasons, static measures such as market share or concentration are unreliable indicators of competitive effects in platform markets, where pricing structures, cross-subsidisation, and non-price competition complicate traditional metrics.

The platform provisions should therefore emphasise limiting principles. Network effects are a potential source of market power, not a presumption of it. Their significance depends on switching costs, multi-homing behaviour, entry conditions, and innovation dynamics. The same logic applies to concerns about potential competition, self-preferencing, interoperability, and data control. Consistent with an error-cost framework, speculative tipping theories should not substitute for an effects-based assessment grounded in evidence that a merger is likely to substantially lessen or prevent competition.

A. Multi-Homing as a Constraint on Platform Power

The Draft Merger Guidelines state the Bureau “may assess the extent of ‘multi-homing,’” noting that multi-homing can allow participants to benefit from network effects while preserving competition among platforms.[73] That observation is important. As ICLE has explained in prior submissions, widespread multi-homing and low switching costs limit the extent to which network effects produce durable market power. Users and business customers can shift activity across platforms, and rivals can reach those users without displacing an incumbent entirely. Many platform environments—including online marketplaces, ride-hailing, and app-based services—exhibit precisely this pattern of overlapping use and ready switching.

The Guidelines should give this principle operational effect. They should identify circumstances in which multi-homing and low switching costs generally weigh against a finding of a substantial lessening or prevention of competition in platform merger cases. Consistent with the need for clear limiting principles, relevant indicators could include evidence that a significant share of participants uses two or more platforms for the same function, that technical and contractual switching costs are low, and that viable alternative channels remain available to reach end users or advertisers.

B. Potential Competition Requires Concrete Evidence

The Draft Merger Guidelines recognise that “dynamic competition may be particularly important in platform markets” and that the “most effective competitive threat” may come from displacement or disintermediation by an innovative or nascent competitor.[74] In innovation-driven markets, the asymmetry of error costs is especially pronounced. False positives that block or burden pro-competitive integrations can chill investment and innovation across the economy, while false negatives are more likely to be corrected over time through entry and technological change. These considerations apply with particular force to platform mergers that may integrate complementary technologies, data, or user groups.

Consistent with an error-cost framework, the Guidelines should require concrete, case-specific evidence that a target is likely to become a substantial competitive constraint on the acquirer. Relevant indicators could include objective signs of expansion, active innovation pipelines, or demonstrated success in a relevant niche. Theoretical possibilities that a complementary or niche service might someday evolve into a rival platform should not, without more, support a finding of a substantial lessening or prevention of competition.

C. Platform-Participant Mergers and Self-Preferencing

For mergers between a platform operator and a participant, the draft states the Bureau “may focus on whether the platform operator would engage in self-preferencing, or otherwise foreclose competing platform participants,” including whether the merged firm could deprive rival platforms of scale or network effects.[75]

As ICLE explained in earlier comments, self-preferencing is typically a form of vertical integration. It often produces pro-competitive benefits, including lower costs, improved quality, stronger investment incentives, and better coordination across complementary services. Empirical research on self-preferencing is mixed and context-specific and does not support categorical presumptions of harm.

The Guidelines should align platform–participant mergers with the general vertical-effects framework. Self-preferencing theories of harm should be pursued only where three conditions are present:

1. the platform possesses durable market power protected by substantial entry or expansion barriers;
2. users face meaningful obstacles to multi-homing or switching; and
3. evidence shows integration is likely to produce net consumer harm—such as higher prices, reduced quality, or diminished innovation—after accounting for efficiencies.

Harm to individual rivals, or a change in their relative position on a platform, should not by itself be treated as harm to competition where those rivals can reach customers through other platforms or distribution channels.

D. Interoperability and Data Theories of Harm Require Evidence

The Draft Merger Guidelines state that mergers involving “tool[s] that make multi-homing easier,” “input[s] or data on which competing platforms rely,” or “complementary, interoperable or related products” may raise concern if they increase incentives to deny interoperability, undermine multi-homing, or raise switching costs.[76] As ICLE explained in earlier comments, broad and open-ended theories of harm in these areas risk turning merger enforcement into de facto regulation of platform design, interoperability, and ecosystem architecture—especially where standards rely on notions of “fairness” or structural preference, rather than demonstrated competitive harm.

The Guidelines should clarify that concerns about interoperability, switching costs, or data control must rest on specific evidence of a profitable and likely foreclosure strategy that would substantially lessen or prevent competition. A general preference for “open” systems or ecosystem neutrality should not substitute for a competitive-effects showing.

The Guidelines should also recognise that not all tools, inputs, or datasets confer market power. Where such assets are reproducible, obtainable from multiple sources, or capable of being developed within a reasonable timeframe, their acquisition will generally not provide the ability to foreclose rivals or entrench market power.

Incorporating these principles—together with the error-cost framework, market-power screens, and the treatment of self-preferencing discussed above—would make the platform-merger section more predictable and more consistent with an effects-based analysis grounded in clear evidence of harm to competition and consumers.

VIII. Labour-Market Effects and Merger Analysis

Bill C-59 authorised the Competition Tribunal to consider labour-market effects in merger review,[77] and the Guidelines appropriately acknowledge that authority. Paragraph 80 explains that, when assessing harm among buyers—including buyers of labour—the Bureau will examine how suppliers respond to the exercise of market power.[78] Footnote 72 properly emphasises the importance of workers’ outside options,[79] and footnote 75 recognises that anticompetitive effects may include reduced wage growth or diminished benefits.[80] These are sound starting points.

The difficulty lies in applying a mirror-image framework to labour and product markets. The draft suggests the Bureau will “apply the same framework to mergers among buyers, including buyers of labour,”[81] yet the economics of monopsony differ in important respects from monopoly analysis. The empirical literature remains unsettled, market definition is unusually challenging, and employment effects cannot be assessed without reference to downstream output and prices. Efficiencies that reduce labour demand may accompany increased production and consumer benefits, making input-side evidence alone insufficient to establish competitive harm.

The final Guidelines should therefore adopt a cautious, evidence-based approach to labour-market enforcement. Labour effects should be evaluated in context, with careful attention to worker mobility, realistic alternatives, and downstream competitive outcomes, rather than through a mechanical application of product-market presumptions.

A. The Evidence on Labour-Market Power Remains Unsettled

The economic literature on labour-market power remains less developed than the industrial-organization literature on product-market competition. As ICLE scholars have explained, many monopsony models in labour economics rely on assumptions about worker mobility, employer conduct, and market structure that have not been tested against the evidentiary standards applied in antitrust cases.[82] Industrial-organization models, by contrast, have been refined through decades of enforcement and litigation, while labour-economics models largely developed without a direct enforcement context.

Direct estimates of labour-market power are also limited and highly variable. Suresh Naidu, Eric Posner, and Glen Weyl report labour-supply elasticity estimates ranging from 0.1 to 4.2, implying workers receive between 9% and 81% of their marginal product, depending on the study.[83] The wide dispersion suggests labour-market conditions differ substantially across occupations and industries, making broad generalisations unreliable.

Empirical studies linking employer concentration to wages face similar methodological constraints. Steven Berry, Martin Gaynor, and Fiona Scott Morton observe that concentration-wage studies “are not ultimately informative about whether monopsony power has grown and is depressing wages.”[84] Ivan Kirov and James Traina likewise find rising markdowns correlate more strongly with technology adoption than with employer concentration, concluding that “concentration is likely an inappropriate measure of labor-market power.”[85]

Given this unsettled evidence, the Bureau should not ground enforcement priorities on indirect and inconclusive measures of competitive harm.

B. Labour-Market Analysis Requires Product-Market Context

The Guidelines assume labour-market analysis parallels product-market analysis. Important conceptual differences limit that analogy. In product markets, competitive harm appears directly in higher prices or reduced output. In labour markets, merger effects cannot be evaluated in isolation from downstream product markets.

Consider a merger that eliminates redundancies and reduces the combined firm’s demand for labour. Viewed only on the input side, fewer hires and lower wages may resemble monopsony power. Yet if the same merger increases downstream output because the combined firm operates more efficiently, the effect is efficiency-enhancing rather than anticompetitive. Ioana Marinescu and Herbert Hovenkamp recognise this point: an efficiency defence in a labour-market case “must show that post-merger reorganization will decrease the need for workers and will not lower total production. Both of these requirements are essential.”[86]

This distinction matters because efficiencies and monopsony can appear identical when assessed solely from employment effects. Consolidating facilities, integrating distribution, or removing overlapping roles may reduce labour demand even where the merger has no anticompetitive purpose or effect. The Bureau cannot distinguish pro-competitive efficiencies from monopsony without examining downstream prices and output.

The consumer-welfare standard compounds this difficulty. A merger that lowers wages but also lowers consumer prices presents a trade-off that competition law does not clearly resolve. Some argue harm to workers alone should establish a violation, regardless of downstream effects.[87] That position sits uneasily with a consumer-focused competition framework. Where consumer benefits and worker impacts diverge, the Guidelines should acknowledge the policy choice rather than assume the answer.

C. Challenges in Defining Labour Markets

Footnote 28 notes the Bureau may need to “adapt our analysis to defining markets for the purchase of a product, including labour.”[88] The challenge is more substantial. Market definition in labour markets raises problems not present in product markets.

Workers are not products. Individuals with similar skills can work across multiple occupations and industries. An accountant may work for a bank, a hotel, or a technology firm. A truck driver may haul goods for a manufacturer, retailer, or logistics company. Because workers can shift across employers and sectors, labour markets are often broader than product markets—not narrower, as the Guidelines elsewhere imply.

Geographic scope presents similar difficulties. Remote work has expanded potential employment areas for many occupations, while commuting patterns vary across regions and job types. The economic literature offers no settled method for defining labour-market boundaries.

These issues matter because market definition determines concentration measures. Narrowly defined labour markets will produce high concentration by construction, regardless of competitive conditions. The Bureau should therefore approach labour-market definitions cautiously and avoid excluding realistic employment alternatives available to workers.

IX. Revisions to the Draft Merger Enforcement Guidelines

This section proposes targeted revisions to improve the Guidelines’ predictability, administrability, and economic coherence. Across multiple topics, the central recommendation is an effects-based approach grounded in clear evidence of substantial competitive harm and informed by the asymmetric costs of enforcement errors.

To that end, the section recommends clarifying that structural presumptions are preliminary screens rather than near-determinative findings; strengthening the treatment of efficiencies and failing-firm scenarios; refining market-definition tools; and adopting workable limiting principles, including market-power screens and safe harbours, particularly for vertical and platform transactions. It also urges careful treatment of labour-market theories and network-effects claims to ensure enforcement rests on demonstrated competitive impairment, not speculative or purely structural concerns.

A. Evidence-Based Enforcement and Restraint in Close Cases

The Guidelines should state, in both the introduction and the executive summary, that the Bureau will pursue enforcement only where supported by evidence of substantial competitive harm. Where competitive effects are ambiguous, disputed, or likely offset by efficiencies, the Bureau should exercise restraint.

This approach follows the error-cost framework. False positives impose greater and more durable costs than false negatives: erroneous intervention chills pro-competitive conduct and deters innovation, while under-enforcement often corrects through entry and competitive pressure. In close cases, non-intervention should therefore be the default.

B. Structural Presumptions and Rebuttal Standards

The Guidelines should clarify that structural presumptions are preliminary screens, not near-determinative findings of harm. Concentration measures correlate poorly with competitive outcomes across many industries. The presumption under subsection 92(2) should trigger closer investigation, not impose a heavy burden on merging parties to prove a negative.

The Bureau should raise, or at minimum not lower, the numerical thresholds. The HHI threshold of 1,800 (with an increase of 100) and the 30% market-share threshold track the 2023 U.S. Merger Guidelines, which agency economists criticised as lacking empirical support. Below an HHI of 2,500, structural presumptions should apply cautiously and be readily rebutted. Below a combined market share of 30%, the Bureau should adopt a strong presumption against enforcement absent extraordinary circumstances, such as coordinated effects, exclusionary conduct, or prevention of entry by a disruptive competitor.

The Guidelines should also specify the evidentiary standard for rebuttal. The draft does not explain what evidence suffices to overcome the presumption on a balance of probabilities. The Bureau should identify categories of rebuttal evidence, including: ease of entry and the likelihood of timely, sufficient entry; expansion by existing competitors; low switching costs or widespread multi-homing; rapid innovation that renders market shares unstable; substantial merger-specific efficiencies; and direct evidence—such as customer testimony, internal documents, or econometric analysis—showing the merger will not substantially lessen competition.

Finally, the Guidelines should clarify how the presumption applies to vertical and conglomerate transactions. Where a transaction is primarily vertical or conglomerate but includes limited horizontal overlap, the presumption should apply only to the horizontal component. Applying horizontal thresholds to vertical relationships lacks economic justification, as discussed in Section VI.

C. Market Definition as a Tool for Assessing Competitive Effects

The Guidelines should emphasise that market definition is a tool for assessing competitive effects, not an independent requirement. Where competitive effects can be analysed directly—through diversion ratios, natural experiments, or bidding data—formal market definition becomes less important. The draft acknowledges this principle but should state it more prominently.

The Bureau should also provide guidance on applying the hypothetical monopolist test to non-price dimensions. The Guidelines recognise that harm may involve quality, variety, privacy, or innovation, yet do not explain how to conduct a SSNIP-equivalent analysis in such settings. In differentiated-product markets, diversion ratios and customer-switching evidence often offer more reliable guidance than formal market boundaries.

Finally, geographic market definition should rest on actual competitive constraints. Relevant considerations include purchasing patterns, transportation costs, and regulatory barriers, rather than administrative or political borders. Where customers routinely source from multiple provinces or countries, those areas should be included in the geographic market.

D. Integrating Efficiencies into Competitive-Effects Analysis

The Guidelines should integrate efficiencies directly into the competitive-effects analysis, rather than treat them as secondary considerations. Marginal-cost reductions bear directly on unilateral effects. If a merger lowers marginal costs sufficiently, the profit-maximising post-merger price may fall even where concentration increases. The current statement that efficiencies are “unlikely to change” the Bureau’s conclusion where significant concerns arise risks nullifying the analysis whenever the structural presumption is triggered.

The Guidelines should also recognise dynamic and fixed-cost efficiencies. Improvements in innovation, product quality, and speed of commercialisation account for a substantial share of consumer-welfare gains in research-intensive industries. Fixed-cost savings affect investment incentives, risk tolerance, and capacity expansion. An approach focused only on short-run price effects will understate these benefits and overstate competitive harm.

Buyer-power efficiencies should not be categorically excluded. Lower input prices obtained through negotiation constitute efficiencies unless they reflect monopsony power that reduces output below competitive levels. Economic theory distinguishes between redistributive transfers and cost reductions that affect downstream pricing incentives.

The Bureau should apply a realistic merger-specificity standard. Merger specificity requires a comparison of practical alternatives that accounts for execution costs, timing, and risk—not a hypothetical scenario in which contractual arrangements costlessly replicate integration. Transaction-cost economics shows firms integrate precisely because contracts cannot replicate certain organisational efficiencies, including protection of relationship-specific investments.

Finally, the Guidelines should clarify how efficiencies interact with the structural presumption. Evidence of substantial, merger-specific efficiencies should be capable of rebutting the subsection 92(2) presumption. Where efficiencies are significant but difficult to quantify precisely, the Bureau should not require a level of certainty greater than that demanded when predicting competitive harm.

E. Ground the Failing-Firm Analysis in Commercial Reality

The Guidelines should clarify that the failing-firm and exiting-assets analysis applies to parts of a business, not only to firm-wide insolvency. Section 93(b) of the Competition Act refers to “the business, or a part of the business, of a party to the merger.” A plant, product line, or local operation that is unprofitable and likely to shut down no longer constrains rivals, even if the parent firm remains solvent. The Guidelines should therefore recognise that exit of a business unit can satisfy the analysis.

The Bureau should also adopt a commercially realistic approach to alternative purchasers. A stringent “no alternative” requirement combined with a mandatory independent search can impose substantial cost and delay on firms already under financial stress. The relevant question is whether the parties conducted a commercially reasonable search in light of their financial position, timing constraints, and market conditions. The competitively preferable purchaser standard should not favour smaller, less-efficient acquirers merely because they hold a lower market share, a concern identified in the economic literature on failing-firm analysis.

Finally, the Guidelines should clarify when liquidation preserves competition. Liquidation promotes competition only where assets are likely to be redeployed within the relevant market. When assets are highly specific, capable buyers are few, or local demand is weak, liquidation may remove those assets from the market entirely. Liquidation should be treated as a superior alternative only where redeployment that maintains competitive constraints is reasonably likely.

F. Clarify the Treatment of Vertical Integration

The Draft Merger Guidelines should delete the unsupported claim that greater vertical integration increases coordination risk. Paragraph 245 states that “the greater the degree of vertical integration in a market, the greater the risk of coordinated behaviour,” yet provides no citation or economic reasoning. Coordinated effects require firms to reach consensus, monitor adherence, and punish deviations—mechanisms vertical integration does not facilitate. If the Bureau maintains this proposition, it should identify supporting evidence.

The Draft Guidelines should establish a vertical safe harbour. Where the merged firm holds less than 40% share at each vertical level, the transaction should be presumed efficient absent specific evidence that foreclosure would competitively impair—not merely disadvantage—rivals.

The draft should recognise the inherent elimination of double marginalisation. Where both vertical levels exercise pricing power, integration removes the double markup by construction. This is a pricing correction, not a speculative post-merger synergy. The guidelines therefore should not apply paragraph 285’s efficiency criteria to this category of gain, and the burden should rest on the Bureau to explain why the effect would not occur, rather than on the parties to prove that it would.

The Bureau should apply consistent treatment of contracts. The Draft Merger Guidelines state that contractual terms are unlikely to eliminate foreclosure harms (footnote 106), yet the guidelines also suggest contracts may replicate integration benefits (footnote 128). Both propositions cannot simultaneously hold. If contracts cannot prevent harms, they likewise cannot fully reproduce integration benefits.

The Bureau should also revise paragraph 288 to provide a meaningful efficiencies pathway. The current statement that efficiency gains are “unlikely to change” conclusions where a merger raises concerns effectively nullifies efficiencies analysis. Parliament retained s. 93(h), which permits the Tribunal to consider “any other factor that is relevant to competition in a market.” Merger-specific efficiencies that enhance consumer welfare fall squarely within that provision, and the Draft Guidelines should give it operational effect.

Finally, the Guidelines should require evidence of competitive impairment, not mere rival disadvantage. Enforcement under the guidelines should target only those vertical mergers where foreclosure would materially impair rivals’ ability to compete and where that impairment would likely harm consumers. A change in competitors’ relative position, standing alone, does not establish a substantial lessening or prevention of competition.

G. Applying Effects-Based Analysis to Platform Mergers

Treat network effects as a factor to be assessed, not a presumption of market power. Network effects exist, but they are often limited and localised. Their competitive significance depends on entry conditions, switching costs, multi-homing, innovation, and the availability of alternatives. Speculative “tipping” theories should not substitute for evidence that a specific merger is likely to substantially lessen or prevent competition.

The Draft Merger Guidelines should also specify when multi-homing and low switching costs defeat a finding of substantial lessening of competition. Relevant evidence could include a significant share of participants using two or more platforms for the same function, low technical and contractual switching costs, and viable alternative channels for reaching users.

The Bureau should also require concrete evidence before concluding that a merger eliminates potential competition. The guidelines should identify objective indicia—such as documented expansion plans, innovation pipelines, or demonstrated success in a relevant segment—showing the target is likely to become a substantial competitive constraint. The mere possibility that a niche or complementary service could evolve into a rival platform should not, without more, support a finding of substantial lessening or prevention of competition.

Finally, the draft should align platform-participant mergers with the vertical-effects framework. Self-preferencing theories of harm should proceed only where three conditions are present: durable market power protected by substantial entry or expansion barriers, meaningful obstacles to multi-homing or switching, and evidence that integration is likely to produce net consumer harm after accounting for efficiencies. Harm to individual rivals, or changes in their relative position on a platform, should not be equated with harm to competition where those rivals can reach customers through other channels.

H. Labour-Market Harms Require Evidence of Downstream Effects

The Draft Merger Guidelines should require evidence of downstream output effects. The Bureau should pursue labour-market theories only where evidence shows a merger reduces output or raises prices in a downstream product market, or where the merger’s predominant effect is wage suppression without offsetting efficiency gains. A merger that reduces labour demand while increasing output is efficiency-enhancing, not anticompetitive.

The Guidelines should apply a demanding evidentiary standard. The economic literature on labour-market power remains less developed than the industrial-organization literature on product-market power, and Canadian legal precedent on labour-market merger effects is limited. The Bureau should require direct evidence of wage suppression or output reduction, rather than reliance on concentration metrics that are especially unreliable in labour markets given cross-industry substitution and variable worker mobility.

The Guidelines should also define labour markets to reflect realistic worker alternatives. Labour-market definitions should account for cross-industry substitutability, commuting patterns, and remote-work opportunities. Artificially narrow markets—limited to specific occupations, bargaining units, or a small set of employers—will overstate concentration and mislead competitive analysis.

Finally, the guidelines should recognise that monopsony analysis differs from monopoly analysis. The assumption that labour-market analysis mirrors product-market analysis obscures the difficulty of distinguishing efficiency-enhancing reductions in labour demand from anticompetitive monopsony. The Bureau cannot draw that distinction without examining the merger’s effects on downstream output and prices.

X. Conclusion

The amendments to the Competition Act enacted through Bills C-56 and C-59 expand the Bureau’s merger-enforcement authority. That authority remains discretionary. These comments address how it can be exercised consistently with the economic evidence on merger effects.

The three concerns identified at the outset—unsupported structural presumptions, the sidelining of efficiencies, and the extension of merger theories into areas with limited evidence—share a common source. Each treats market structure as a reliable proxy for competitive outcomes and discounts the mechanisms through which mergers benefit consumers. The industrial-organization literature reviewed here does not support that approach. Concentration measures correlate weakly with competitive harm across industries and market conditions. Vertical mergers typically benefit consumers. Efficiencies—particularly the elimination of double marginalisation—directly affect pricing and output in ways structural indicators cannot capture.

Section II’s error-cost framework gives these findings practical significance. Blocking a merger that would benefit consumers imposes durable costs: forgone efficiencies, reduced investment, and chilled innovation. By contrast, if a harmful merger escapes challenge, entry, expansion, and technological change often erode supra-competitive returns over time. This asymmetry does not favour inaction. It favours precision-targeted intervention where evidence of harm is substantial, and restraint where evidence is ambiguous or offsetting benefits are likely.

The recommendations in Section IX would move the Guidelines toward that standard. Structural presumptions should operate as screening tools rather than near-determinative findings. Efficiencies should form part of the competitive-effects analysis. Vertical-merger enforcement should focus on demonstrable competitive impairment, not structural suspicion. Platform and labour-market theories should require concrete evidence of consumer harm. These changes would reduce false positives without weakening enforcement against genuinely anticompetitive mergers.

ICLE appreciates the Bureau’s openness to public input through this consultation and remains available to provide further analysis on any of the issues addressed in these comments.

[1] Competition Bureau, Proposed Merger Enforcement Guidelines (Nov. 13, 2025), https://competition-bureau.canada.ca/en/how-we-foster-competition/consultations/proposed-merger-enforcement-guidelines [hereinafter Draft Merger Guidelines].

[2] Geoffrey Manne, Dirk Auer, Daniel Gilman & Lazar Radic, Comments from the International Center of Law and Economics on the Future of Competition Policy in Canada (Mar. 31, 2023), https://laweconcenter.org/wp-content/uploads/2023/03/ICLE-Canada-Comments.pdf

[3] Ian Adams, Eric Fruits, Brian Albrecht, Daniel J. Gilman & Geoffrey A. Manne, Comments of the International Center for Law & Economics to the Competition Bureau Canada: Anti-Competitive Conduct and Agreements (Jan. 27, 2026), https://laweconcenter.org/wp-content/uploads/2026/01/Competition-Bureau-Canada-Comments-2026.pdf (Appendix AA) [hereinafter ACCA Comments].

[4] Geoffrey A. Manne, Ian Adams, Brian Albrecht, Dirk Auer, Lazar Radic & Mario A. Zúñiga, Comments of the International Center for Law & Economics: Canada Competition Bureau’s Update of the Merger Enforcement Guidelines (Jan. 12, 2025), https://laweconcenter.org/wp-content/uploads/2025/01/Canada-Merger-Reform-Comments-January-2025.pdf (Appendix BB).

[5] An Act to Amend the Excise Tax Act and the Competition Act, S.C. 2023, c. 31, pt. 2 [hereinafter Bill C-56]; Fall Economic Statement Implementation Act, 2023, S.C. 2024, c. 15, pt. 5, div. 6 [hereinafter Bill C-59].

[6] Brian Albrecht, Competition Increases Concentration, Truth on the Mkt. (Aug. 16, 2023), https://truthonthemarket.com/2023/08/16/competition-increases-concentration.

[7] See Brian Albrecht, What Competition Scholars Should Know About the 2025 Economics Nobel, Truth on the Mkt. (Oct. 14, 2025), https://truthonthemarket.com/2025/10/14/what-competition-scholars-should-know-about-the-2025-economics-nobel (“Market structure is an outcome of this competitive process, not just a cause of competitive behavior.”).

[8] Chad Syverson, Macroeconomics and Market Power: Context, Implications, and Open Questions, 33 J. Econ. Persps. 23, 34 (2019).

[9] See Harold Demsetz, Industry Structure, Market Rivalry, and Public Policy, 16 J.L. & Econ. 1 (1973); see also, e.g., Richard Schmalensee, Inter-Industry Studies of Structure and Performance, in 2 Handbook of Industrial Organization 951–1009 (Richard Schmalensee & Robert Willig eds., 1989); William N. Evans, Luke M. Froeb & Gregory J. Werden, Endogeneity in the Concentration-Price Relationship: Causes, Consequences, and Cures, 41 J. Indus. Econ. 431 (1993); Steven Berry, Market Structure and Competition, Redux, FTC Micro Conference (Nov. 2017), https://www.ftc.gov/system/files/documents/public_events/1208143/22_-_steven_berry_keynote.pdf; Nathan Miller et al., On the Misuse of Regressions of Price on the HHI in Merger Review, 10 J. Antitrust Enf’t 248 (2022).

[10] Harold Demsetz, The Intensity and Dimensionality of Competition, in Harold Demsetz, The Economics of the Business Firm: Seven Critical Commentaries 137, 140-41 (1995).

[11] Harold Demsetz, Industry Structure, Market Rivalry, and Public Policy, 16 J.L. & Econ. 1 (1973); see also Harold Demsetz, The Market Concentration Doctrine: An Examination of Evidence and a Discussion of Policy (AEI–Hoover Policy Studies No. 7, 1973), https://masonlec.org/site/rte_uploads/files/GAI/Readings/Economics%20Institute/Demsetz_Market%20Concentration%20Doctrine.pdf.

^[12] Steven Berry, Martin Gaynor & Fiona Scott Morton, Do Increasing Markups Matter? Lessons from Empirical Industrial Organization, 33 J. Econ. Persp. 44, 48 (2019); see also Jonathan Baker & Timothy F. Bresnahan, Economic Evidence in Antitrust: Defining Markets and Measuring Market Power, John M. Olin Program in L. & Econ., Stanford Law Sch. Working Paper 24 (Sept. 2006) (“structure-conduct-performance empirical methods have largely been discarded in economics.”).

[13] Dennis W. Carlton, Market Definition: Use and Abuse, 3 Competition Pol’y Int’l 3 (2007), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=987061.

[14] Christopher Garmon, The Accuracy of Hospital Screening Methods, 48 Rand J. E Econ. 1068, 1070 (2017) (Examining post-merger price changes across 28 hospital mergers and first issued as an FTC Bureau of Economics working paper.).

[15] William J. Baumol, John C. Panzar & Robert D. Willig, Contestable Markets and the Theory of Industry Structure 222 (rev. ed. 1988), available at https://archive.org/details/contestablemarke0000baum_e1g8 (“[F]reedom of entry, indeed the mere threat of incursions by entrants into the market, may effectively discipline the monopolist, even if entry is never successful. It can force the monopolist to curb his avarice and forgo profits he might otherwise have enjoyed. Indeed, in the absence of entry barriers, in perfectly contestable markets, it can force him to accept earnings no higher than those available under perfect com¬ petition. Potential competition can also force the monopolist to produce with maximal efficiency, and to hunt down and utilize fully every opportunity for innovation.”); see also David J. Teece, Understanding Dynamic Competition: New Perspectives on Potential Competition, “Monopoly,” and Market Power, Working Paper (May 22, 2025), https://ssrn.com/abstract=5356023.

[16] Baumol, Panzar & Willig, Id.

[17] Org. for Econ. Coop. & Dev., Competition in the Pharmaceutical Industry, OECD Doc. DAFFE/CLP(2000)29 (Feb. 6, 2001) (“For many products which face few rivals in their therapeutic class, the primary competitive threat is the threat that rival firms will develop substitutes.”).

[18] Stephen J. Nickell, Competition and Corporate Performance, 104 J. Pol. Econ. 724 (1996).

[19] Richard Blundell, Rachel Griffith & John Van Reenen, Market Share, Market Value and Innovation in a Panel of British Manufacturing Firms, 66 Rev. Econ. Stud. 529 (1999).

[20] U.S. Dep’t of Justice & Fed. Trade Comm’n, Merger Guidelines 2, 5–6 (Dec. 18, 2023), https://www.ftc.gov/system/files/ftc_gov/pdf/2023_merger_guidelines_final_12.18.2023.pdf.

[21] U.S. Dep’t of Justice & Fed. Trade Comm’n, Horizontal Merger Guidelines 19 (Aug. 19, 2010), https://www.justice.gov/atr/file/810276/dl?inline.

[22] Id.

[23] Expediency can be seen as a motivating factor not only in the weakly justified new thresholds, but in a novel “presumption of illegality” absent from prior U.S. merger guidelines—and notably avoided in the Competition Bureau’s proposal. See, e.g., Daniel J. Gilman, What Do We Do with Presumptions in Antitrust?, Truth on the Mkt. (Jan. 25, 2024), https://truthonthemarket.com/2024/01/25/what-do-we-do-with-presumptions-in-antitrust.

[24] Nathan Miller et al., On the Misuse of Regressions of Price on the HHI in Merger Review, 10 J. Antitrust Enf’t 248 (2022), https://doi.org/10.1093/jaenfo/jnac009.

[25] John Asker, Kostis Hatzitaskos, Bob Majure, Ana McDowall, Nathan Miller & Aviv Nevo, Comments on the January 2022 DOJ and FTC RFI on Merger Enforcement, FTC-2022-0003-1847 (Apr. 20, 2022), https://www.regulations.gov/comment/FTC-2022-0003-1847.

[26] Geoffrey A. Manne, Dirk Auer, Brian Albrecht, Eric Fruits, Daniel J. Gilman & Lazar Radi?, Comments of the International Center for Law & Economics on the FTC & DOJ Draft Merger Guidelines, Docket No. FTC-2023-0043-0001 (Sept. 18, 2023), https://laweconcenter.org/wp-content/uploads/2023/09/ICLE-Draft-Merger-Guidelines-Comments-1.pdf.

[27] Asker et al., supra note 25, at 5. For an additional critique of the structuralist push by Biden administration enforcers, see Bruce H. Kobayashi & Timothy J. Muris, Turning Back the Clock: Structural Presumptions in Merger Analyses and Revised Merger Guidelines, Competitive Enter. Inst. Study (Feb. 22, 2023), https://cei.org/studies/turning-back-the-clock-structural-presumptions-in-merger-analyses-and-revised-merger-guidelines.

[28] 2003 FCA 53, leave to appeal refused, [2003] S.C.C.A. No. 421.

[29] Competition Act § 93(h).

[30] Draft Merger Guidelines ¶ 286.

[31] Draft Merger Guidelines ¶ 288.

[32] See, e.g., Dennis W. Carlton & Jeffrey M. Perloff, Modern Industrial Organization (2nd ed.1994) 36-37 (“reasons for mergers and acquisitions” include efficiencies—greater optimal scale, scope synergies, and improved management.).

[33] Appendix BB 5.

[34] See, e.g., Joseph Farrell & Carl Shapiro, Horizontal Mergers: An Equilibrium Analysis, 80 Am. Econ. Rev. 107, 114–16 (1990) (“But if the merger also generates synergies, the new firm’s marginal costs may be lower than those of either of its constituent firms, and this cost reduction may be enough to cause the price to fall.”)

[35] Gregory Sidak & David J. Teece, Dynamic Competition in Antitrust Law, 5 J. Competition L. & Econ. 581 (2009).

[36] Oliver E. Williamson, Economies as an Antitrust Defense: The Welfare Tradeoffs, 58 Am. Econ. Rev. 18 (1968) (“[I]n the occasional case where efficiency and market power consequences exist, can economies be dismissed on the grounds that market power effects invariably dominate? If they cannot, then a rational treatment of the merger question requires that an effort be made to [estimate the tradeoffs].”)

[37] Org. for Econ. Coop. & Dev., Monopsony and Buyer Power, OECD Doc. DAF/COMP(2008)38 (2009), https://www.oecd.org/content/dam/oecd/en/publications/reports/2009/12/monopsony-and-buyer-power_cf60bdd1/36a2b824-en.pdf.

[38] Compare Lars-Hendrik Röller, Johan Stennek & Frank Verboven, Efficiency Gains from Mergers, in European Merger Control: Do We Need an Efficiency Defence? (Fabienne Ilzkovitz & Roderick Meiklejohn eds., 2006) (“Real cost-savings are those savings that correspond to some savings of productive resources in the economy. Rationalisation, economies of scale, technological progress, and slack reduction are all real cost-savings. Also some purchasing economies are real cost-savings.”), with Org. for Econ. Coop. & Dev., Dynamic Efficiencies in Merger Analysis, OECD Doc. DAF/COMP(2007)41 (May 15, 2008), https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/05/dynamic-efficiencies-in-merger-analysis_bffeeed6/df6017f9-en.pdf (“[A] post-merger rise in price leads to a redistribution of income from buyers, who pay the higher prices, to the company and its shareholders, who receive the higher profit. Since the dollars of welfare loss suffered by the buyers are balanced by the dollars of welfare gain accruing to the producers, many economists have treated this merely as a transfer of income between the two groups, in which case society as a whole is unaffected… distributional changes generally should not be considered to affect social welfare and, therefore, that such changes can be ignored.”).

[39] Id.

[40] Peter G. Klein & João Fernando Rossi Mazzoni, The Make-or-Buy Decision Revisited, in Handbook of New Institutional Economics 447 (Claude Ménard & Mary M. Shirley eds., 2d ed. 2025).

[41] See, e.g., Oliver E. Williamson, The Vertical Integration of Production: Market Failure Considerations, 61 Am. Econ. Rev. 112 (1971) (explaining that, under “bounded rationality” and “transactional failures,” market contracts cannot match a firm’s internal control and incentive structures). See also Oliver E. Williamson, Transaction-Cost Economics: The Governance of Contractual Relations, 22 J.L. & Econ. 233 (1979) (explaining that significant transaction costs—driven by bounded rationality and opportunism—keep market contracts from matching the efficiencies of internal organisation).

[42] Jeffrey T. Macher & Barak D. Richman, Transaction Cost Economics: An Assessment of Empirical Research in the Social Sciences, 10 Bus. & Pol. 1 (2008) (review of empirical studies finds that, as transactions grow more complex or require relationship-specific investment, firms shift from simple contracts to integrated governance, because contracts cannot match a firm’s administrative control and flexibility).

[43] Draft Merger Guidelines ¶ 277.

[44] Draft Merger Guidelines ¶ 277.

[45] Draft Merger Guidelines ¶ 280.

[46] Draft Merger Guidelines ¶ 281.

[47] Org. for Econ. Coop. & Dev., Competition Comm., The Failing Firm Defence, OECD Doc. DAF/COMP(2009)38, at 12 (Aug. 17, 2010), https://www.oecd.org/content/dam/oecd/en/publications/reports/2010/10/the-failing-firm-defence_a0e37e7e/c90c3d1e-en.pdf (“In some instances, the merger under review involves the acquisition of a firm’s division. In those cases, the merging parties may argue that the exit of that particular division from the market would occur (i) whether or not the merger materialises and (ii) irrespective of the financial health of the parent company.”).

[48] Draft Merger Guidelines ¶ 282.

[49] Draft Merger Guidelines ¶ 279.

[50] Draft Merger Guidelines ¶ 283.

[51] Id.

[52] Lars Persson, The Failing Firm Defense, 53 J. Indus. Econ. 175 (2005).

[53] Draft Merger Guidelines ¶ 283.b.

[54] Draft Merger Guidelines ¶ 283.c.

[55] Draft Merger Guidelines ¶ 224.

[56] Draft Merger Guidelines ¶ 230.

[57] Francine Lafontaine & Margaret E. Slade, Exclusive Contracts and Vertical Restraints: Empirical Evidence and Public Policy, in Handbook of Antitrust Economics 391 (Paolo Buccirossi ed., 2008); see also Francine Lafontaine & Margaret E. Slade, Transaction Cost Economics and Vertical Market Restrictions—Evidence, 55 Antitrust Bull. 587 (2010); Geoffrey Manne, Kristian Stout & Eric Fruits, The Fatal Economic Flaws of the Contemporary Campaign against Vertical Integration, 68 U. Kan. L. Rev. 923 (Jun. 2020).

[58] Francine Lafontaine & Margaret E. Slade, Vertical Integration and Firm Boundaries: The Evidence, 45 J. Econ. Literature 629, 680 (2007).

[59] Global Antitrust Inst., Vertical Mergers, Comment on the Fed. Trade Comm’n’s Hearings on Competition and Consumer Protection in the 21st Century 6–8 (Sept. 6, 2018), https://gaidigitalcommons.org/vertical-mergers.

[60] Chiara Farronato, Andrey Fradkin & Alexander MacKay, Self-Preferencing at Amazon: Evidence from Search Rankings, Nat’l Bureau Econ. Rsch., Working Paper No. 32,359, at 3–4 (2025).

[61] Benjamin Klein & Kevin M. Murphy, Vertical Restraints as Contract Enforcement Mechanisms, 31 J.L. & Econ. 265 (1988).

[62] Draft Merger Guidelines ¶ 241(a) (“[G]enerally speaking, the greater the degree of vertical integration in a market, the greater the risk of coordinated behaviour.”).

[63] Daniel P. Reiffen & Michael Vita, Comment: Is There New Thinking on Vertical Mergers?, 63 Antitrust L.J. 917, 920-21 (1995).

[64] Michael A. Salinger, Vertical Mergers and Market Foreclosure, 103 Q.J. Econ. 345, 346-48 (1988).

[65] Draft Merger Guidelines ¶ 229.

[66] Draft Merger Guidelines n.128 (“Firms may claim that non-horizontal mergers allow the merged firm to ‘internalize’ existing double marginalization. We will consider these claims in light of the criteria in paragraph 285, including considering whether any potential elimination of double marginalization is merger-specific, or could be partially or entirely achieved through other contracts, agreements or supply relationships.”).

[67] Draft Merger Guidelines n.128 (“Firms may claim that non-horizontal mergers allow the merged firm to ‘internalize’ existing double marginalization. We will consider these claims in light of the criteria in paragraph 285, including considering whether any potential elimination of double marginalization is merger-specific, or could be partially or entirely achieved through other contracts, agreements or supply relationships.”).

[68] Draft Merger Guidelines n.106 (“Where a merger otherwise results in the ability and incentive to foreclose, we presume that contractual terms are unlikely to eliminate all potential means of foreclosure, unless the contrary is clearly demonstrated.”).

[69] Draft Merger Guidelines ¶ 288.

[70] Competition Act, R.S.C. 1985, c. C-34, s. 93(h).

[71] Draft Merger Guidelines § 4.4.3.

[72] Draft Merger Guidelines ¶ 248.

[73] Draft Merger Guidelines ¶ 249.

[74] Draft Merger Guidelines ¶ 250.

[75] Draft Merger Guidelines ¶ 251.b.

[76] Draft Merger Guidelines ¶¶ 251c, 252.

[77] Fall Economic Statement Implementation Act, 2023, S.C. 2024, c. 15, s. 255 (adding subsection 93(1.1) to the Competition Act, allowing the Tribunal to consider “harm to competition in a labour market” when assessing whether a merger substantially lessens or prevents competition).

[78] Draft Merger Guidelines ¶ 80(b).

[79] Draft Merger Guidelines n.72 (“If the merged firm controls a large share of purchases in the relevant market, a presumption of substantial harm may apply and we are likely to investigate further. For example, we may consider the alternatives available to suppliers (or workers) and whether there are barriers that prevent new buyers or employers from entering or expanding.”).

[80] Draft Merger Guidelines n.75.

[81] Draft Merger Guidelines n.89.

[82] See Brian Albrecht, Dirk Auer & Geoffrey A. Manne, Labor Monopsony and Antitrust Enforcement: A Cautionary Tale, Int’l Ctr. for L. & Econ. (May 2025), https://laweconcenter.org/resources/labor-monopsony-and-antitrust-enforcement-a-cautionary-tale (“The models commonly employed in labor economics to support these claims rely on assumptions about worker mobility, employer conduct, and market structure that likely oversimplify real-world dynamics…. The economic models commonly used to study labor markets have not been subjected to the same level of antitrust scrutiny as those employed in industrial-organization (IO) economics to analyze product markets.”).

[83] Suresh Naidu, Eric A. Posner & E. Glen Weyl, Antitrust Remedies for Labor Market Power, 132 Harv. L. Rev. 536, 549-50 (2018) (summarising studies reporting residual labour elasticities ranging from 0.1 to 4.2).

[84] Steven Berry, Martin Gaynor & Fiona Scott Morton, Do Increasing Markups Matter? Lessons from Empirical Industrial Organization, 33 J. Econ. Persps. 44, 62 (2019).

[85] Ivan Kirov & James Traina, Labor Market Power and Technological Change in U.S. Manufacturing 3, 32, Working Paper (2023).

[86] Ioana Marinescu & Herbert Hovenkamp, Anticompetitive Mergers in Labor Markets, 94 Ind. L.J. 1031, 1062 (2019).

[87] See C. Scott Hemphill & Nancy L. Rose, Mergers that Harm Sellers, 127 Yale L.J. 2078, 2083 (2018) (arguing that “harm to input markets suffices to establish an antitrust violation”).

[88] Draft Merger Guidelines n.28.

ICLE Director of Law & Economics Programs Gus Hurwitz participated in a panel at the 2026 Silicon Flatirons Flagship Conference on “The Administrative State and Tech Access,” alongside Blake E. Reid, Jennifer Huddleston, Chris Lewis, and Tejas Narechania. The discussion examined how recent Supreme Court decisions, executive actions, and shifting federal-state and independent-agency relationships affect technology policy and access, including implications for the Universal Service Fund, the Federal Trade Commission, the Humphrey’s Executor precedent, and the Major Questions Doctrine. Video of the full panel is embedded below.

Executive Summary

Generative artificial intelligence (AI) has diffused with unusual speed since late 2022. By late 2024, nearly 40% of U.S. adults ages 18–64 reported using AI tools, a pace that exceeds comparable stages for personal computers and the internet. That rapid uptake has sharpened two policy questions: whether AI will generate measurable gains in output and productivity at the aggregate level, and whether the adjustment process will produce labor-market disruption large enough to justify new regulatory intervention.

The empirical literature points to a clear pattern. Controlled field experiments and randomized trials document large productivity gains at the task and firm level, often alongside quality improvements. Across writing, customer support, software development, accounting, law, and translation, studies report 15% to more than 50% reductions in task-completion time, meaningful quality gains, and disproportionately large benefits for less-experienced workers, producing “skill compression” within occupations.

At the same time, aggregate labor-market indicators through 2024–2025 show limited disruption, despite rapid adoption. Most datasets find little evidence of economywide job loss or wage decline. Where effects appear, they are concentrated in entry-level segments of highly exposed occupations, while senior employment remains largely stable. Adjustment to date has occurred through task reallocation and within-firm productivity gains, rather than mass displacement.

Macroeconomic projections remain uncertain. Credible estimates range from modest productivity gains to large output increases, depending on assumptions about task coverage, diffusion speed, organizational redesign, and complementary investment. Disagreement in the macro literature reflects divergent assumptions and measurement challenges, including the well-documented productivity J-curve, rather than conflicting data.

The evidence also informs competition policy. AI lowers minimum viable scale in many downstream markets, facilitating entry by startups and small firms, even as upstream model development may remain concentrated. Open-source components and contracting already provide broad access to AI inputs, weakening claims that mandatory access or data-sharing rules are necessary to preserve competition.

Overall, the literature offers limited support for regulatory approaches premised on widespread worker displacement, durable monopoly power, or exclusionary “data moats.” A more defensible approach emphasizes targeted enforcement of existing law, reduced regulatory fragmentation, and investment in complements such as skills, governance, and infrastructure.

The issue brief concludes with an extensive annotated bibliography summarizing the empirical and theoretical studies underlying these findings.

I. Macroeconomic Effects of AI: Evidence, Assumptions, and Uncertainty

The macroeconomic literature on AI remains driven largely by models, rather than realized output data. That imbalance is partly mechanical. Large-scale diffusion began only recently, and firm-level reorganization and supply-chain adjustment typically lag initial adoption.

Two additional issues shape how existing macro evidence should be interpreted: measurement and complements.

Diane Coyle and John Lourenze S. Poquiz (2025) argue that standard national accounts may understate AI’s economic contribution when gains appear as quality improvements, time savings, or outputs that markets price poorly. When workers produce higher-quality text, code, or designs in less time, welfare gains at the firm level can exceed what revenue-based aggregates capture. Measurement limits do not imply the absence of output gains, but they weaken inferences drawn from a short window of aggregate data.

Erik Brynjolfsson, Daniel Rock, and Chad Syverson’s (2021) “productivity J-curve” highlights a second interpretive challenge. General-purpose technologies can raise long-run productivity even as measured productivity initially stagnates or falls.[1] Early adoption requires firms to invest in organizational redesign, worker training, and process change before efficiency gains appear in output statistics. The J-curve implies a transition period in which complementary investments determine whether productivity gains ultimately materialize.

Observed macroeconomic flatness therefore supports competing interpretations. One view holds that AI delivers modest aggregate gains even after diffusion, either because the affected task share of GDP is limited or because current capabilities remain poorly matched to high-value work. Another view holds that the economy remains in an adjustment phase, with organizational and measurement frictions delaying visible gains. The literature does not yet resolve which interpretation dominates, but it provides a structured framework for evaluating for future research.

A. Assumption-Driven Divergence in Macroeconomic Forecasts of AI

The sharpest disagreements in the macroeconomic literature turn on three margins: task share, adoption, and complementary investment.

Daron Acemoglu (2025) develops a task-based macro model that distinguishes “easy-to-learn” tasks with objective verification from “hard-to-learn” tasks requiring context-sensitive judgment. He applies Hulten’s Theorem to bound aggregate gains by the GDP share of tasks that AI can perform.[2] Under his parameterization, AI raises total factor productivity by less than 0.66% over 10 years. The estimate is deliberately conservative and mechanism-driven: when AI excels only on a subset of tasks and struggles where verification is costly or ambiguous, aggregate gains remain limited even with broad diffusion.

Alex Arnon and Kent Smetters (2025), writing for the Penn Wharton Budget Model, offer a middle-ground estimate grounded in exposure classifications, adoption assumptions, and micro-level evidence. They conclude that roughly 10% of current U.S. GDP could be affected in the short run, rising to 15% over two decades under partial adoption. Their calculations assume average labor-cost savings of about 25% from current tools, potentially increasing to 40% as systems improve. The results hinge on adoption speed and on whether measured cost reductions translate into higher output rather than rent reallocation.

More optimistic projections assume faster diffusion and stronger downstream effects. Joseph Briggs and Devesh Kodnani (2023) estimate a 7% increase in annual global GDP over 10 years—roughly $7 trillion—based on broad occupational task exposure. Erkan Erdem and Dileep Birur (2025), using a dynamic computable general equilibrium framework, estimate that rapid adoption could raise U.S. GDP by about $2.48 trillion by 2030.[3] The Congressional Research Service synthesizes these projections and notes that estimates are typically positive but highly sensitive to diffusion timelines that may unfold over decades, as with earlier general-purpose technologies (Lida Weinstock and Paul Tierno, 2025).

Taken together, these forecasts differ less over data than over assumptions. Aggregate effects depend on three factors: diffusion across sectors, whether productivity gains expand output rather than reshuffle rents, and whether organizational redesign occurs at scale. The wide range of estimates reflects divergence along these margins rather than measurement error alone.

Isabel Aldasoro, Sebastian Doerr, Leonardo Gambacorta, and Daniel Rees (2024) underscore this point by modeling AI adoption as a positive productivity shock. They show that aggregate output, consumption, and investment rise in both the short and long run, while inflation dynamics depend on expectations. When households and firms anticipate productivity gains, demand can increase early and push inflation upward. When gains arrive unanticipated, supply may initially outpace demand, producing disinflation.

For policy, the implication is straightforward: short-run stabilization outcomes during AI adoption can diverge sharply from long-run productivity effects. Regulators should not infer long-run welfare consequences from near-term inflation responses alone.

B. Plausible AI-Driven Geographic Divergence, with Uncertain Causal Attribution

The Council of Economic Advisers (2026) argues that AI could generate an international growth divergence analogous to earlier industrial transitions, with the United States positioned to benefit from advantages in investment, compute capacity, and innovation ecosystems. The report points to stronger growth among so-called “Pax Silica” economies and the United States’ large share of global compute resources. It reflects an official policy position favoring rapid deployment and diffusion.

The empirical case for attributing recent growth differentials to AI remains limited. Short observation windows and confounding macroeconomic forces complicate efforts to isolate AI’s contribution from broader cyclical, fiscal, and geopolitical factors.

Shahid Yusuf’s (2025) United Nations Development Programme analysis similarly highlights the risk of divergence but shifts the focus to distributional effects. It emphasizes uneven gains within countries and across regions, particularly in Asia-Pacific economies facing long-run productivity slowdowns.

II. Labor-Market Effects of AI: Evidence of Adjustment Rather Than Displacement

Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock (2024) provide a foundational mapping between occupational tasks and large language model capabilities. They estimate that roughly 80% of U.S. workers have at least 10% of their tasks exposed to LLM assistance, and about 19% have 50% or more of tasks exposed. They stress that “exposure” measures technical feasibility for task assistance, not a prediction of job destruction. That distinction matters, because task reallocation and mass job elimination carry fundamentally different welfare implications.

Xianguo Huang (2025) extends the exposure framework across countries and demographic groups. He finds higher exposure in higher-income economies and disproportionate exposure among women and mid-education workers in some contexts. Consistent with the task-based view, the analysis characterizes current AI systems as primarily augmenting human labor, rather than automating it outright.

A. Limited Aggregate Labor-Market Disruption Through 2024–2025

Several studies report null or modest aggregate labor-market effects despite rapid AI adoption. The Budget Lab at Yale finds no clear relationship between AI exposure and unemployment through August 2025 (Martha Gimbel et al., 2025). Anders Humlum and Emilie Vestergaard (2025) link survey-reported ChatGPT use to Danish administrative records across 11 exposed occupations and find essentially zero effects on earnings or hours through 2024.

U.S.-based evidence points in the same direction. Jonathan S. Hartley, Filip Jolevski, Vitor Melo, and Brendan Moore (2026) report that 35.9% of U.S. workers used generative AI by December 2025 and find small positive wage effects, with no statistically significant declines in job openings or employment in exposed occupations. Bharat Chandar’s (2025) Current Population Survey analysis similarly finds no aggregate employment decline, while documenting heterogeneity across education levels and occupations.

Taken together, these studies find no evidence of immediate economywide labor displacement through 2024–2025. The results instead point to early adjustment through task reallocation, quality improvement, and within-firm productivity gains, rather than broad-based layoffs.

B. Entry-Level Effects in Some AI-Exposed Occupations

While aggregate indicators remain stable, several studies identify concentrated entry-level effects in highly exposed segments. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen (2025), using ADP payroll data, estimate that workers ages 22–25 in highly exposed occupations experienced employment declines of roughly 16% relative to trend following ChatGPT’s release, while senior employment remained stable. Bouke Klein Teeselink (2025), examining the United Kingdom, finds that exposed firms reduced employment and hiring concentrated among junior and entry-level roles and posted lower advertised salaries for exposed occupations.

These findings point to a plausible adjustment channel. AI can automate discrete tasks that have traditionally served as entry-level work, reducing marginal demand for junior labor without displacing more senior workers.

The evidence suggests structural adjustment in career ladders and human-capital formation, rather than broad job loss. If entry-level pathways narrow, labor markets may respond by shifting screening and training mechanisms or reallocating junior workers toward tasks that remain complementary to AI. The literature has not yet resolved whether these effects persist or fade as firms adapt.

C. Complementarity and Skill Compression in Many AI-Exposed Settings

Andrew C. Johnston and Christos A. Makridis (2025) find that higher-exposure sectors experienced wage and employment gains, particularly among younger and more educated workers, while roles characterized by direct substitution saw declines. The pattern points to task-level complementarity, rather than economywide labor replacement.

Evidence from controlled experiments shows that realized productivity gains also depend on workers’ ability to judge when AI assistance improves outcomes. Andrew Caplin et al. (2024), studying an age-classification task, find that AI raises performance across ability levels but reduces performance dispersion most when users are well calibrated about their own skills. Low-ability participants who accurately recognized their limitations achieved the largest gains by relying selectively on AI, while overconfident or underconfident users underperformed relative to calibrated peers. A counterfactual exercise suggests that universal calibration would nearly double AI’s inequality-reducing effects.

Across settings, micro-level evidence points to skill compression. AI tools disproportionately boost output among lower-performing workers, narrowing performance gaps within job categories. This pattern recurs across productivity studies and carries distributional implications: AI can reduce inequality within occupations, even as it reshapes demand for certain entry-level roles.

III. Within-Firm Productivity Gains and Their Limits

The micro-level evidence provides the strongest empirical support in the current literature. Shakked Noy and Whitney Zhang (2023) conduct a randomized experiment with 453 professionals and find that ChatGPT use reduces task-completion time by roughly 40% and increases output quality by about 18%, with larger gains concentrated among initially lower-performing workers. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond (2025) examine a Fortune 500 customer-support deployment and find a 15% average productivity increase, measured as issues resolved per hour, and a 36% increase for workers in the bottom skill quintile. Customer satisfaction remained stable, while attrition among newer workers declined sharply. Together, these results show that AI can transmit best practices and raise performance at the lower end of the skill distribution.

Field evidence from professional services points in the same direction. Jung Ho Choi and Chloe L. Xie (2025) analyze accounting work at a technology firm and survey 277 accountants, finding that AI adoption correlates with an 18% increase in weekly client support and a reallocation of roughly 9% of work hours from routine data entry to higher-value tasks, such as client communication. AI use reduced monthly book-closing timelines by 7.5 days and increased ledger detail by 12%, indicating quality improvements alongside time savings. The authors also document complementarity between professional expertise and AI confidence scores: experienced accountants used model outputs to target review effort, rather than replace judgment. A framed experiment reveals occasional overreliance on inaccurate suggestions, underscoring the importance of verification protocols.

Evidence from legal practice reinforces these findings. Daniel Schwarcz et al. (2025) conduct a randomized controlled trial evaluating AI tools used by law students on complex legal tasks. Both a retrieval-augmented legal system and a general-purpose reasoning model substantially improved document quality across clarity, organization, and analytical depth. Students using AI completed assignments 50% to 130% faster than control groups. The study also identifies functional differentiation across tools: retrieval-augmented systems reduced citation errors, while reasoning models improved substantive analysis. These results suggest that limitations observed in earlier model generations are not fixed and that appropriate tool selection and workflow design can mitigate risks without restricting deployment.

Software development studies show similarly large effects. Sida Peng et al. (2023) report that GitHub Copilot users complete coding tasks 55.8% faster in controlled settings, with larger gains among less experienced developers. Kevin Zheyuan Cui et al. (2025), studying nearly 5,000 developers across three large field experiments, find a 26.08% increase in weekly task completion, driven by higher adoption and disproportionately larger gains for junior developers. These findings undermine claims that AI primarily benefits top performers and instead show AI reducing frictions for early-career workers.

Experimental evidence from translation extends the pattern. Ali Merali (2024) conducts a randomized trial with 300 professional translators and links increased training compute to economic outcomes. A tenfold increase in compute reduced completion time by 12.3%, improved quality by 0.18 standard deviations, and raised earnings per minute by 16.1%. Lower-skilled translators experienced gains roughly four times larger than their higher-skilled counterparts.

The magnitude and replication of these effects across contexts suggest that productivity gains arise consistently in specific task categories—including writing, customer support, software development, and translation—while remaining sensitive to task structure and workflow integration.

A. The Jagged Technological Frontier and Deployment Constraints

Fabrizio Dell’Acqua et al. (2023) describe a “jagged technological frontier,” in which AI exhibits uneven capabilities across tasks that appear similar in difficulty. Within this frontier, AI can substantially improve productivity and quality for some complex tasks, while producing errors on others that seem straightforward. The unevenness requires knowledge workers to exercise judgment in deciding when AI assistance is appropriate and when human oversight remains essential.

In a Boston Consulting Group field experiment using GPT-4, Dell’Acqua et al. (2023) find that AI improved performance on tasks within its capability boundary but reduced performance on tasks just beyond it. The decline stemmed from overreliance on plausible but incorrect model outputs. The results show that productivity effects depend critically on task selection and verification systems, not on model capability alone.

This literature identifies an internal governance challenge for firms. Verification protocols, worker training, and clear task assignment can mitigate boundary failures more effectively than broad constraints on model development. A blanket regulatory approach that raises adoption costs across all use cases would fail to target the source of error and risk suppressing productivity gains where AI performs reliably.

B. Why Intra-Firm Productivity Gains Do Not Automatically Translate to Macro Growth

The micro-level evidence indicates substantial productivity potential, but its translation to macroeconomic outcomes depends on complementary investments. The J-curve framework predicts that firms incur adjustment costs—through reorganization, training, and process redesign—before realizing large productivity gains (Erik Brynjolfsson, Daniel Rock, and Chad Syverson, 2021).

Measurement frictions can further delay visibility in aggregate data. Diane Coyle and John Lourenze S. Poquiz (2025) show that standard GDP statistics often miss quality improvements and time savings, weakening inferences drawn from short-run macro indicators.

Diffusion also remains uneven. Some occupations exhibit large productivity effects in controlled experiments, while others show limited exposure or face high verification costs (Tyna Eloundou et al., 2024; Daron Acemoglu, 2025). Together, these constraints explain why macroeconomic effects may lag or appear smaller than micro-level gains would suggest.

IV. Entrepreneurship and Business Formation in the AI Economy

Junhui Jeff Cai et al. (2025) analyze administrative business-formation data using a difference-in-differences design around ChatGPT’s release. They find increased entry by first-time and resource-constrained founders and show that post-ChatGPT firms tend to have fewer shareholders and smaller founding teams. The mechanism is direct: AI substitutes for managerial, operational, and technical tasks that previously required additional hires or cofounders. By lowering the minimum viable team size, AI reduces entry costs, increases the number of entrants, and strengthens downstream competition. These effects arise independently of concentration at the foundation-model layer.

Survey evidence from the Organisation for Economic Co-operation and Development (2025b) reinforces this pattern. About 31% of small and medium-sized enterprises across seven countries had adopted generative AI by 2024, and 83% of adopters reported no change in staffing levels. Firms cited reduced workloads and improved performance and many described AI adoption as a response to labor shortages, rather than a labor-replacement strategy. Non-adopters most often cited lack of suitability (57%), legal or data-privacy concerns (54%), and insufficient internal skills (50%).

These findings carry two policy-relevant implications. First, uncertainty about compliance and legal risk remains a material barrier to adoption, suggesting that clearer rules and regulatory harmonization could accelerate diffusion. Second, workforce capability constrains uptake, pointing toward training and institutional learning, rather than command-and-control regulation aimed at model developers.

Capital-market evidence highlights a complementary margin. Iuri Struta (2024) reports more than $20 billion invested in generative AI startups through Q3 2024. Abu Bakkar Siddik, Yong Li, and Anna Min Du (2024), studying 556 generative AI startups, find that investor influence—measured by the number of investors, lead investors, and funding rounds—strongly predicts total funding, while measures of “technological influence,” such as IT spending and patenting, do not. Access to capital networks and distribution channels remains central to entry and scaling, underscoring the importance of contract freedom and partnership formation for competitive AI markets.

V. Dynamic Competition and Market Structure in AI Markets

The Organisation for Economic Co-operation and Development’s (2025a) analysis of competitive dynamics in downstream markets concludes that AI can lower entry barriers and the minimum efficient scale by automating functions that previously required large, specialized teams. This finding aligns with the entrepreneurship evidence. The result is a plausible decoupling: upstream model development may exhibit concentration driven by fixed costs in compute and training, while downstream markets experience increased entry, experimentation, and rivalry.

OECD competition roundtables also document mixed effects from vertical integration between model providers and application firms (OECD, 2025a). Some forms of integration may foreclose rivals, while others improve coordination, reduce transaction costs, and accelerate diffusion. Absent evidence of durable foreclosure, vertical contracting is generally presumptively efficiency-enhancing. Broad structural presumptions risk deterring procompetitive arrangements that facilitate deployment and innovation.

Cross-country evidence underscores the role of human capital, infrastructure, and trade openness. Alessandra Bonfiglioli et al. (2025) find that countries with larger STEM graduate pipelines, higher internet penetration, and greater export volumes hold comparative advantages in AI-intensive industries, while restrictive digital trade policies correlate with weaker export performance. Alex Haag’s (2025) Federal Reserve analysis identifies U.S. advantages in infrastructure, compute capacity, and investment conditions, alongside constraints facing China in advanced semiconductors and persistent gaps in cloud scale and private investment across Europe. These findings suggest that national competitiveness turns on enabling inputs and that heavy restrictions on cross-border data flows or digital trade can impose meaningful costs.

Evidence on model openness further complicates simple competition narratives. Thibault Schrepel and Jason Potts (2025) evaluate 11 foundation models using an 18-variable index focused on licensing and governance. They show that openness operates along a spectrum, rather than a binary divide, with many models clustered in the middle and only modest score differences between systems often labeled “open” or “closed.” For competition policy, the implication is narrow but important: arguments that hinge on categorical labels should be tested against enforceable license terms and governance structures.

Open-source AI adoption reinforces this point. Anna Hermansen and Cailean Osborne (2025) report that 89% of organizations using AI incorporate open-source components, with higher adoption rates among small and mid-sized firms due to lower deployment costs and greater flexibility. Survey and case evidence indicate that open-source systems often reduce business-unit costs by more than 50%, by enabling inter-organizational collaboration and faster development cycles. In specialized domains such as health care, open-source models perform comparably to proprietary alternatives, while offering superior integration flexibility in manufacturing and edge-computing environments. These findings suggest that access to AI inputs frequently emerges through market mechanisms, rather than regulation, and that restrictions on contracting or licensing risk disrupting existing diffusion pathways.

Policy commentary from the International Center for Law & Economics (ICLE) echoes these concerns. In comments responding to U.S. Justice Department (DOJ) proposals, Geoffrey A. Manne et al. (2024) argue that mandatory data-sharing and access requirements can reduce competition by weakening investment incentives and undermining startups’ ability to form partnerships that provide compute and distribution. They further contend that data advantages are not necessarily exclusionary because performance gains from data can saturate, shifting competition toward algorithmic and product innovation.

Giorgio Castiglia’s (2025) case for dynamic competition policy reinforces this perspective. In fast-moving technology markets, innovation—not static market share—defines rivalry. In AI markets, where performance improvements can rapidly reallocate competitive advantage, competition enforcement that treats current positions as durable risks misdiagnosing the competitive process.

VI. Policy Implications: Strategic Forbearance and Complementary Investment

ICLE’s comments to the U.S. Office of Science and Technology Policy advance a case for “strategic forbearance,” urging regulators to rely on existing technology-neutral law while agencies modernize rules built around assumptions of human operators or static systems (Eric Fruits, Ben Sperry, and Kristian Stout. 2025). The argument rests on uncertainty and rapid technological change. When model capabilities, best practices, and deployment structures evolve quickly, prescriptive ex ante rules risk imposing high compliance costs, while failing to target actual harm channels. Kristian Stout’s later comments similarly warn against burdensome frameworks and emphasize the difficulty of defining “AI” in legally stable terms as technology evolves (Stout, 2025a). Overbroad definitions can impose compliance costs across low-risk uses, creating entry barriers unrelated to the harms regulators seek to address.

Stout (2025b) further argues that federal preemption of conflicting state AI regulations would reduce market fragmentation and compliance costs, facilitating interstate deployment and lowering fixed costs for smaller firms. From a competition perspective, fragmentation functions as an entry barrier. Legal overhead scales with the number of jurisdictions, rather than with output, disproportionately burdening startups and small and medium-sized enterprises.

The broader literature supports applying existing legal frameworks to AI rather than constructing sector-wide regulatory regimes. Many concerns raised in policy debates map cleanly onto established doctrines. Consumer deception falls within consumer protection and unfair practices law. Employment discrimination remains governed by civil rights and labor statutes. Product defects and safety risks are addressed through product-liability rules and sector-specific safety regulation. This approach aligns with the economic principle of targeting: regulation should address specific externalities or market failures, rather than impose general constraints that burden benign uses. It also avoids raising diffusion costs in contexts where the literature finds productivity gains and improvements in work quality (Shakked Noy and Whitney Zhang, 2023; Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, 2025).

Where genuine regulatory gaps exist, Fruits, Sperry, and Stout (2025) recommend pilot programs, waivers, and conditional approvals that allow agencies to learn about risks and benefits before imposing permanent requirements. This approach preserves flexibility while generating evidence.

Multiple strands of evidence underscore the importance of complements. Micro-level experiments show that workers gain substantially when trained to use AI tools effectively, with especially large benefits for less-experienced workers (Noy and Zhang, 2023; Brynjolfsson, Li, and Raymond, 2025; Kevin Zheyuan Cui et al., 2025; Jung Ho Choi and Chloe L. Xie, 2025; Daniel Schwarcz et al., 2025). Evidence on the jagged technological frontier shows that poor task assignment and overreliance can reduce performance, strengthening the case for internal governance and verification, rather than technology-wide restrictions (Fabrizio Dell’Acqua et al., 2023). Comparative-advantage studies highlight the role of STEM supply and infrastructure and link restrictive digital-trade policies to weaker performance in AI-intensive industries (Alessandra Bonfiglioli et al., 2025).

Policy that raises adoption costs through compliance burdens risks delaying the complement-building phase described by the productivity J-curve (Erik Brynjolfsson, Daniel Rock, and Chad Syverson, 2021). Taken together, the evidence points toward an enabling posture: reduce fragmentation, clarify rules, and invest in the complements that allow productivity gains to materialize.

VII. Conclusion

The empirical literature supports several conclusions with relatively high confidence. First, controlled workplace studies consistently show large productivity gains from AI in task categories such as writing, customer support, software development, and translation, often improving quality as well as speed (Shakked Noy and Whitney Zhang, 2023; Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, 2025; Sida Peng et al., 2023; Kevin Zheyuan Cui et al., 2025; Ali Merali, 2024). These gains appear repeatedly across firms, occupations, and experimental designs and are strongest among initially lower-performing workers, producing skill compression, rather than elite-only benefits.

Second, aggregate labor-market effects through 2024–2025 remain limited in most datasets. Studies using administrative records and large surveys find little evidence of economywide job loss or wage decline despite rapid adoption (Martha Gimbel et al., 2025; Anders Humlum and Emilie Vestergaard, 2025; Jonathan S. Hartley et al., 2026). At the same time, several datasets identify pressure in entry-level segments of highly exposed occupations, particularly among younger workers and new hires (Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, 2025; Bouke Klein Teeselink, 2025). The emerging pattern is adjustment at the margin—through task reallocation and changes in career ladders—rather than broad displacement.

Third, macroeconomic effects remain uncertain. Credible estimates range from modest productivity gains to large output increases, depending on assumptions about task share, diffusion speed, and complementary investment (Daron Acemoglu, 2025; Alex Arnon and Kent Smetters, 2025; Joseph Briggs and Devesh Kodnani, 2023; Erkan Erdem and Dileep Birur, 2025). Measurement challenges and the productivity J-curve further complicate interpretation, as organizational redesign and intangible investment can delay visible gains in aggregate data (Erik Brynjolfsson, Daniel Rock, and Chad Syverson, 2021; Diane Coyle and John Lourenze S. Poquiz, 2025). Divergence across forecasts reflects disagreement over these margins, not simple data error.

Several cross-cutting themes emerge. AI’s economic impact depends less on raw model capability than on deployment context, governance, and complements. The “jagged technological frontier” shows that productivity gains hinge on task selection and verification, not blanket automation (Fabrizio Dell’Acqua et al., 2023). Diffusion appears strongest where firms invest in training, workflow redesign, and internal controls. At the market level, AI lowers minimum viable scale and facilitates entry downstream, even as upstream model development may remain concentrated due to fixed costs. Competition, in this setting, is dynamic and innovation-driven, rather than static and share-based.

For policy, these findings point toward restraint coupled with focus. After two years of rapid adoption, the most defensible posture remains strategic forbearance: targeted enforcement of existing law, combined with efforts to reduce regulatory fragmentation and support complementary investment in skills, governance, and infrastructure (Eric Fruits, Ben Sperry, and Kristian Stout, 2025; Kristian Stout, 2025b). Consumer protection, civil rights, and product-liability doctrines already address many identified risks. Where uncertainty persists, pilot programs, waivers, and conditional approvals offer a way to learn without locking in premature mandates.

Proposals for forced data sharing, mandatory access, or structural restrictions on partnerships require stronger evidence of persistent market failure and clearer proof that intervention improves welfare net of dynamic costs (Geoffrey A. Manne et al., 2024). The current literature provides less support for those premises than the speed of AI adoption might suggest. The central policy challenge is not to slow diffusion, but to ensure that institutions, skills, and governance evolve quickly enough for productivity gains to materialize broadly and sustainably.

Annotated Bibliography

• Acemoglu, Daron, The Simple Macroeconomics of AI, 40 Pol’y 13 (2025).

Develops a task-based macroeconomic model to estimate AI’s aggregate productivity effects. The model distinguishes easy-to-learn tasks, which have objective success metrics, from hard-to-learn tasks, which require contextual judgment. Finds that generative AI will raise total factor productivity by less than 0.66% over a 10-year horizon. Current systems perform well on easy tasks but deliver diminishing returns on hard tasks, which account for a larger share of overall economic activity. Applies Hulten’s Theorem to show that aggregate productivity gains are bounded by the GDP share of tasks affected by AI. As a result, even substantial improvements at the task level translate into modest economywide effects. Offers the most conservative credible estimate in the current literature and pinpoints specific technical constraints that limit AI’s macroeconomic impact. The analysis is essential for assessing the plausible range of AI-driven growth forecasts.

• Aldasoro, Iñaki, Sebastian Doerr, Leonardo Gambacorta & Daniel Rees, The Impact of Artificial Intelligence on Output and Inflation, BIS Working Paper No. 1179, Bank for Int’l Settlements (Apr. 2024), https://www.bis.org/publ/work1179.pdf.

Employs a multi-sector macroeconomic model showing that AI adoption functions as a positive productivity shock that raises aggregate output, consumption, and investment in both the short and long run. Simulations show that inflation effects hinge on expectations. When households and firms anticipate productivity gains, immediate demand growth produces inflation. When gains arrive unanticipated, supply initially outpaces demand, resulting in disinflation. At the sectoral level, the authors find little correlation between an industry’s initial AI exposure and its long-term output growth. The results also indicate that AI adoption in consumption-good sectors generates substantially larger aggregate output gains than adoption in investment-good sectors, reflecting amplification through sectoral production linkages.

• Arnon, Alex & Kent Smetters, The Projected Impact of Generative AI on Future Productivity Growth, Penn Wharton Budget Model (Sept. 8, 2025), https://budgetmodel.wharton.upenn.edu/issues/2025/9/8/projected-impact-of-generative-ai-on-future-productivity-growth.

Presents a detailed report combining task-exposure data, adoption assumptions, and empirical productivity studies to estimate AI’s macroeconomic effects. The analysis concludes that roughly 10% of current U.S. GDP could be affected in the short run, rising to 15% over two decades under partial-adoption scenarios. Assumes average 25% labor-cost savings from currently available AI tools, with potential gains reaching 40% as the technology matures. Builds a sector-level model that incorporates workforce composition and task-automation potential to translate firm-level efficiencies into aggregate outcomes. Synthesizes evidence from multiple micro-level productivity studies, including Brynjolfsson, Noy & Zhang, and Peng et al., to project economywide effects. Positions its estimates between conservative and highly optimistic forecasts and provides a policy-relevant baseline for evaluating long-term budget and growth implications associated with AI adoption.

• Bick, Alexander, Adam Blandin & David J. Deming, The Rapid Adoption of Generative AI, NBER Working Paper No. 32966 (2025), https://www.nber.org/papers/w32966.

Documents that generative AI adoption reached nearly 40% of the U.S. population ages 18–64 by late 2024, with diffusion occurring faster than for personal computers or the internet. Analyzes demographic adoption patterns and finds parallels to early PC uptake. Uses Real-Time Population Survey (RPS) data to measure adoption frequency and intensity across users. Demonstrates rapid technology uptake and estimates potential productivity gains based on reported time savings, rather than on market-deregulation effects.

• Bonfiglioli, Alessandra, Rosario Crinò, Mattia Filomena & Gino Gancia, Comparative Advantage in AI-Intensive Industries: Evidence from U.S. Imports, CESifo Working Paper No. 11642 (2025), https://ssrn.com/abstract=5116412.

Investigates the determinants of global competitiveness in AI-intensive industries using a dataset of U.S. imports from 68 countries across 79 industries from 1999 to 2019. Constructs a novel AI-intensity index based on occupations that require machine-learning and data-analysis skills to measure industry exposure. Finds that countries with larger supplies of STEM graduates, broader internet penetration, and higher overall export volumes exhibit a strong comparative advantage in AI-intensive sectors. Links these structural factors to stronger export performance in high-technology industries. Shows that restrictive digital-trade regulations—particularly those affecting infrastructure and cross-border data flows—correlate with significantly lower export performance in AI-intensive fields. The results remain robust across multiple controls and instrumental-variable analyses using historical scientific data. Highlights the central role of human capital, digital infrastructure, and open regulatory environments in shaping comparative advantage in the digital economy.

• Briggs, Joseph & Devesh Kodnani, The Potentially Large Effects of Artificial Intelligence on Economic Growth, Goldman Sachs Econ. Rsch. (Mar. 26, 2023), https://www.gspublishing.com/content/research/en/reports/2023/03/27/d64e052b-0f6e-45d7-967b-d7be35fabd16.pdf.

Estimates that generative AI could raise annual global GDP by 7%—nearly $7 trillion—over a 10-year period. Combines data on the task content of more than 900 occupations with adoption-rate assumptions to forecast productivity gains. Projects that roughly 300 million full-time jobs could face automation exposure, while noting that historical patterns show worker displacement often offset by new job creation. Frames the results as a baseline scenario with “potentially large macroeconomic effects” contingent on adoption timelines.

• Brynjolfsson, Erik, Bharat Chandar & Ruyu Chen, Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence, Digit. Econ. Lab (Nov. 2025), https://digitaleconomy.stanford.edu/app/uploads/2025/11/CanariesintheCoalMine_Nov25.pdf.

Analyzes ADP payroll data covering millions of workers to identify AI’s employment effects. Uses Poisson-regression event-study estimations to control for firm-level shocks and isolate the relationship between AI exposure and hiring patterns following ChatGPT’s release. Finds that workers ages 22–25 in highly AI-exposed occupations experienced a 16% employment decline relative to trend, while senior-level employment remained stable. Documents that employment effects concentrate in occupations where AI automates, rather than augments, labor. Provides early large-scale evidence of AI’s differential workforce impacts, identifying entry-level workers as disproportionately affected.

• Brynjolfsson, Erik, Danielle Li & Lindsey Raymond, Generative AI at Work, 140 J. Econ. 889 (2025).

Analyzes data from 5,172 customer-support agents at a Fortune 500 enterprise to evaluate deployment of a GPT-based conversational assistant. Uses a staggered-rollout design to identify causal effects, supplemented by a pilot randomized controlled trial. Finds a 15% average productivity increase, measured as issues resolved per hour, with a 36% gain among agents in the bottom skill quintile. The system diffused best practices from top performers, effectively delivering real-time coaching at scale. Reports stable customer-satisfaction scores alongside faster resolution times. Attrition declined by roughly 10 percentage points (40%) among newer agents with AI access. Demonstrates that generative AI can raise the productivity floor while improving work experience, challenging predictions that automation inevitably degrades working conditions.

• Brynjolfsson, Erik, Daniel Rock & Chad Syverson, The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, 13 Econ. J.: Macroeconomics 333 (2021).

Explains why general-purpose technologies often fail to raise measured productivity in the near term despite clear technical capabilities. Firms must undertake complementary investments in “organizational restructuring,” worker training, and process redesign before realizing efficiency gains. Treats these intangible investments as short-run costs that temporarily depress measured productivity. Historical evidence from electrification and information technology shows that productivity gains emerge only after substantial lags, in some cases taking “a generation” or longer. Frames this dynamic as the upward trajectory of a productivity J-curve. The framework suggests the economy may remain in the investment phase of the curve with respect to AI, meaning the absence of immediate productivity surges does not undermine more optimistic long-term projections.

• Cai, Junhui Jeff, Xian Gu, Liugang Sheng, Mengjia Xia, Linda Zhao & Wu Zhu, AI as “Co-founder”: GenAI for Entrepreneurship, arXiv Preprint arXiv:2512.06506 (2025), https://arxiv.org/pdf/2512.06506.

Uses administrative firm-formation data and a difference-in-differences design centered on ChatGPT’s November 2022 release to assess effects on business creation. Finds that generative AI facilitated market entry by first-time founders and resource-constrained entrepreneurs, particularly in industries downstream of AI capabilities. Reports that firms formed after ChatGPT’s introduction had fewer shareholders and smaller founding teams, consistent with AI substituting for managerial, technical, and operational labor at the startup stage. Interprets these patterns as evidence that AI supplies domain knowledge and functional capabilities that previously required larger teams. Shows that generative AI can lower barriers to entrepreneurship beyond automating existing tasks by enabling new business models viable at smaller scale. Highlights implications for market structure and business dynamism.

• Caplin, Andrew, David J. Deming, Shangwen Li, Daniel J. Martin, Philip Marx, Ben Weidmann & Kadachi Jiada Ye, The ABC’s of Who Benefits from Working with AI: Ability, Beliefs, and Calibration, NBER Working Paper No. 33021 (Oct. 2024), http://www.nber.org/papers/w33021.

Presents a controlled experiment using an age-classification task to examine how individual ability and belief calibration—defined as the accuracy of self-assessment—shape AI-driven productivity gains. Finds that AI assistance improves performance across users but produces the largest effects when individuals are well calibrated in assessing their own skills. Shows that low-ability participants who accurately recognize their limitations realize the greatest performance gains because they rely more effectively on AI recommendations. In contrast, miscalibration—whether overconfidence or underconfidence—produces suboptimal decision-making and constrains the technology’s benefits. Counterfactual analysis indicates that if all users held perfectly calibrated beliefs about their abilities, AI’s equalizing effects on performance would nearly double. Highlights calibration training as a key workforce complement to AI deployment.

• Castiglia, Giorgio, Rethinking Antitrust: The Case for Dynamic Competition Policy, Tech. & Innovation Found. (Oct. 2025), https://www2.itif.org/2025-dynamic-competition.pdf.

Argues that innovation-based competition should take priority in technology markets where creative destruction can rapidly displace incumbent firms. Although not AI-specific, the framework maps directly onto foundation-model markets, where algorithmic advances can quickly reconfigure competitive positions. Critiques antitrust approaches that prioritize static market-share metrics over dynamic innovation incentives. Contends that such frameworks risk misidentifying competitive harm in fast-moving, technology-driven sectors. Provides a theoretical basis for skepticism toward structural interventions that could preserve existing market configurations at the expense of future innovation. Offers a policy-relevant lens for evaluating competition proposals in AI markets.

• Chandar, Bharat, Tracking Employment Changes in AI-Exposed Jobs, Digit. Econ. Lab (Aug. 1, 2025), https://ssrn.com/abstract=5384519.

Analyzes U.S. Current Population Survey (CPS) data from late 2022 through early 2025 to assess generative AI’s labor-market effects. Finds no aggregate decline in employment or earnings among occupations with the highest AI exposure, countering broad displacement claims. Disaggregated results show substantial heterogeneity. High-exposure, college-degree roles—such as software development—experienced employment growth, while lower-education roles, including customer service, recorded employment declines. Identifies a divergence between online job-postings data and realized employment outcomes. While postings signaled a downturn in tech hiring, observed employment levels remained stable, suggesting postings may serve as an unreliable proxy for labor-market conditions.

• Choi, Jung Ho & Chloe L. Xie, Human + AI in Accounting: Early Evidence from the Field, Univ. & Mass. Inst. of Tech. (Sept. 2025), https://ssrn.com/abstract=5240924.

Analyzes field data from a technology firm alongside survey responses from 277 accountants to assess workplace AI adoption. Finds that AI use correlates with an 18% increase in weekly client support and a reallocation of roughly 9% of work hours from routine data entry to higher-value tasks, including business communication. Documents measurable quality improvements in financial reporting. AI adoption reduced monthly book-closing timelines by 7.5 days and increased ledger-account detail by 12%. Identifies a “Human + AI” complementarity in which experienced professionals use AI confidence scores to target review efforts. A framed field experiment also finds that accountants sometimes over-rely on inaccurate AI outputs, underscoring the continuing importance of professional expertise as a control mechanism.

• Council of Econ. Advisers, Artificial Intelligence and the Great Divergence, Off. of the President of the U.S. (2026), https://www.whitehouse.gov/wp-content/uploads/2026/01/Artificial-Intelligence-and-the-Great-Divergence-5.pdf.

Posits that AI could produce a second “Great Divergence” analogous to the Industrial Revolution, with technologically leading nations accelerating ahead of peers. Reports that so-called “Pax Silica” economies—the United States and key allies—averaged 2.5% real GDP growth through Q3 2025, compared with 1.1% across G7 nations. Documents U.S. leadership in core AI inputs, including investment flows, compute capacity—estimated at 74% of the global total—and development of large-scale AI systems. Links this position to national innovation capacity and geopolitical advantage. Frames federal policy around accelerating deployment through infrastructure expansion and deregulation. Projects that exponential capability gains—measured by performance metrics doubling within months—will translate into sustained productivity growth. Reflects the Council of Economic Advisers’ position favoring rapid diffusion and regulatory restraint to preserve U.S. technological leadership.

• Coyle, Diane & John Lourenze S. Poquiz, Making AI Count: The Next Measurement Frontier, NBER Working Paper No. 34330 (Oct. 2025), http://www.nber.org/papers/w34330.

Argues that conventional GDP statistics understate AI’s economic impact by failing to capture quality improvements and intangible-capital formation. When AI enables workers to produce higher-quality outputs in less time, input- and revenue-based measures miss a substantial share of resulting welfare gains. Examines both conceptual and practical challenges in measuring AI-driven productivity improvements. Proposes extensions to national accounting frameworks designed to better capture AI-related value creation. Explains that the absence of a pronounced productivity surge in official statistics may reflect measurement constraints rather than a lack of real economic gains. Provides an interpretive framework for assessing macroeconomic data in periods of rapid technological change.

• Cui, Kevin Zheyuan, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng & Tobias Salz, The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers, Working Paper (Aug. 2025), https://ssrn.com/abstract=4945566.

Investigates generative AI’s productivity effects on high-skilled labor through three large-scale field experiments involving nearly 5,000 software developers at Microsoft, Accenture, and a Fortune 100 firm. Evaluates deployment of an AI-based coding assistant across real-world production environments. Finds that developers with AI access increased weekly task completion by 26.08%, alongside measurable gains in code updates and compilations. Reports heterogeneous effects across experience levels, with junior and less-experienced developers showing higher adoption rates and larger productivity improvements than senior peers. Indicates that generative AI can materially increase output in high-skilled occupations without degrading work quality, with particularly strong effects among early-career professionals.

• Dell’Acqua, Fabrizio, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon & Karim R. Lakhani, Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, Harv. Bus. Sch. Working Paper No. 24-013 (2023), https://www.hbs.edu/ris/Publication%20Files/24-013_d9b45b68-9e74-42d6-a1c6-c72fb70c7282.pdf.

Reports results from a field experiment with Boston Consulting Group consultants evaluating GPT-4’s effects on productivity and work quality. Finds substantial performance gains on tasks within the system’s capability range, alongside performance declines on tasks just beyond that frontier. Attributes these declines to overreliance on AI-generated suggestions for problems the system could not reliably solve. Introduces the “jagged frontier” concept to describe uneven AI capability boundaries, where superficially similar tasks can differ sharply in automation potential. Demonstrates the importance of human oversight and the risks of automation complacency. Concludes that effective deployment requires workers to exercise judgment about when to trust, verify, or override AI outputs, highlighting organizational and managerial challenges in integrating these systems.

• Eloundou, Tyna, Sam Manning, Pamela Mishkin & Daniel Rock, GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, 384 Science 1306 (2024).

Develops a foundational task-exposure classification for large language models. Uses expert human raters alongside GPT-4 to evaluate which occupational tasks AI systems could perform. Finds that roughly 80% of U.S. workers have at least 10% of their tasks exposed to AI capabilities. Exposure skews toward higher-income occupations, particularly those involving programming and writing, while roles centered on science and critical-thinking tasks show negative correlations. Estimates that about 19% of workers could face exposure across 50% or more of their tasks. Emphasizes that “exposure” reflects technical feasibility for augmenting labor efficiency, not a prediction of automation or displacement. Establishes an empirical baseline for assessing AI’s economic and policy implications.

• Erdem, Erkan & Dileep Birur, Generative AI and Economic Growth: A New Approach to Measuring Its Potential Economic Impact, KPMG (Nov. 2025), https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2025/gen-ai-economic-growth.pdf.

Employs a dynamic computable general-equilibrium (CGE) model to forecast the economic effects of generative AI adoption across the U.S. and global economies from 2024 through 2050. Models three scenarios—rapid adoption, slow adoption, and a no-generative-AI baseline—to assess how integration rates and workforce-upskilling strategies would shape GDP, wages, and employment across labor categories. Projects that generative AI could generate trillions of dollars in incremental economic value. Under rapid-adoption assumptions, U.S. GDP could increase by approximately $2.48 trillion by 2030. Finds that realizing these gains depends on sustained investment in labor productivity and strategic workforce development to manage transitional labor-market adjustment costs.

• Fruits, Eric, Ben Sperry & Kristian Stout, Comments of the International Center for Law & Economics: Office of Science and Technology Policy RFI, Regulatory Reform on Artificial Intelligence, Int’l Ctr. for L. & Econ. (Oct. 27, 2025), https://laweconcenter.org/wp-content/uploads/2025/10/OSTP-AI-Comments-2025.pdf.

Advocates a policy framework of “strategic forbearance” for artificial intelligence governance. Argues that because AI functions as a rapidly evolving general-purpose technology, policymakers should prioritize existing technology-neutral statutes—including those addressing fraud, discrimination, and product safety—rather than adopt premature, prescriptive mandates that risk constraining innovation. Identifies structural barriers to AI deployment. Highlights “regulatory mismatch,” in which legacy federal rules assume human operators or static devices, and points to a fragmented landscape of conflicting state-level laws that impose elevated compliance costs. Recommends that federal agencies deploy administrative tools—such as waivers, pilot programs, and conditional approvals—to generate empirical evidence and modernize outdated regulatory frameworks, while preserving innovation incentives.

• Joshua Gans, The Microeconomics of Artificial Intelligence (MIT Press, 2025).

Presents a book-length analytical framework that conceptualizes AI as a technology that reduces prediction costs. Develops formal models showing that, as prediction becomes cheaper, the value of complementary human judgment rises—particularly in setting objectives, preferences, and values that guide how predictions are used. Analyzes AI’s effects on pricing, organizational design, and market structure across a range of applications. Applies a microeconomic lens to policy domains including competition law, privacy, and misinformation. Argues that AI improves decision-making efficiency while introducing risks tied to identifiable market failures, including bias and externalities. Provides a theoretical foundation for regulatory approaches that target these discrete failures, rather than imposing broad constraints on capability development.

• Gimbel, Martha, Molly Kinder, Joshua Kendall & Maddie Lee, Evaluating the Impact of AI on the Labor Market: Current State of Affairs, Budget Lab at Yale (Oct. 1, 2025), https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs.

Assesses AI’s labor-market effects using industry- and occupation-level data through August 2025. Examines whether high-exposure sectors exhibit shifts in occupational composition or elevated unemployment rates. Finds no clear correlation between AI exposure and aggregate unemployment. The null results point to broad labor-market resilience during the current adoption phase. Complements studies that identify localized or demographic-specific impacts by showing overall employment stability. Indicates that effects observed in narrower populations have not yet translated into economywide disruption and may take longer to materialize.

• Haag, Alex, The State of AI Competition in Advanced Economies, FEDS Notes (Oct. 6, 2025), https://www.federalreserve.gov/econres/notes/feds-notes/the-state-of-ai-competition-in-advanced-economies-20251006.html.

Presents a Federal Reserve comparative analysis of AI competition across advanced economies, focusing on investment flows, research output, and enabling infrastructure. Finds that the United States retains structural advantages in compute capacity, digital infrastructure, and investment conditions. China has expanded research output and accelerated adoption but faces constraints in advanced semiconductor manufacturing and segments of the software ecosystem. European economies lag in private investment and hyperscale cloud capacity. Concludes that existing initial conditions favor continued U.S. leadership, even as international competition intensifies. Supports policy approaches that reinforce foundational enablers, including chip supply, STEM education, and digital infrastructure.

• Hartley, Jonathan S., Filip Jolevski, Vitor Melo & Brendan Moore, The Labor Market Effects of Generative Artificial Intelligence, Working Paper (Jan. 2026), https://ssrn.com/abstract=5136877.

Presents a comprehensive survey of U.S. workers assessing generative AI adoption and labor-market outcomes. Finds that 35.9% of workers reported using generative AI tools by December 2025, with adoption concentrated among younger, college-educated, and higher-earning employees. Identifies small but statistically significant positive wage effects alongside no measurable change in job openings or aggregate employment across AI-exposed occupations, despite rapid uptake. Links survey responses—aggregated at the occupation level—with administrative data to measure relationships among AI exposure, earnings, and labor demand. Provides one of the most current snapshots of AI diffusion and near-term labor-market effects. Indicates that predictions of immediate, widespread displacement have not materialized and situates current adjustment dynamics within an early adoption phase.

• Hawk, Ryan, Jeroen van Hoof, Nicki Wakefield & Allen Webb, The Leader’s Guide to Value in Motion, PwC (Apr. 29, 2025), https://www.pwc.com/leaders-guide-value-in-motion.

Uses general-equilibrium modeling to quantify the economic effects of artificial intelligence and climate change on global markets through 2035. Projects that AI-driven productivity gains could raise global GDP by nearly 15%, offsetting an estimated 7% contraction tied to physical climate risks and roughly 3% in losses from assets stranded by decarbonization. Maps current industrial sectors to future growth-domain value pools, estimating that these reconfigured markets could generate more than $132 trillion by 2035. Outlines three forward-looking growth scenarios, ranging from 24.9% to 37.2% cumulative expansion, showing how long-term performance will depend on technological adoption, institutional trust, and geopolitical stability.

• Hermansen, Anna, & Cailean Osborne, The Economic and Workforce Impacts of Open Source AI: Insights from Industry, Academia, and Open Source Research Publications, Linux Found. (May 2025), https://www.linuxfoundation.org/hubfs/Research%20Reports/lfr_marketimpacts25_052725a.pdf.

Analyzes the economic and workforce effects of open-source artificial intelligence (OSAI), identifying it as a key driver of industry innovation and cost efficiency. Reports that 89% of organizations deploying AI incorporate open-source components, with smaller firms adopting OSAI at higher rates due to accessibility and lower deployment costs. Finds that OSAI adoption can reduce business-unit costs by more than 50%, driven by interorganizational collaboration and accelerated development cycles for high-quality models. Shows that open-source models in specialized sectors, including health care, perform comparably to proprietary systems, while offering greater flexibility for operational integration in manufacturing and edge-computing environments.

• Huang, Xianguo, Labor Market Exposure to AI: From GenAI to Future AGI, Working Paper WP/25-12, ASEAN+3 Macroeconomic Rsch. Off. (Nov. 2025), https://amro-asia.org/wp-content/uploads/2025/11/GenAI_Labour_Huang2025_20251107.pdf.

Finds that higher-income economies face greater labor-market exposure to generative AI due to higher concentrations of cognitive and digital-infrastructure-dependent roles. Concludes that current systems primarily augment human productivity, rather than fully automate tasks, while noting that this balance could shift toward displacement with the emergence of artificial general intelligence. Presents a case study of Brunei showing moderate aggregate exposure but uneven sectoral effects. Finance, insurance, and administrative services display the highest transformation risk. Identifies demographic disparities in exposure. Women and workers with midlevel educational attainment face disproportionately higher exposure relative to other workforce groups.

• Humlum, Anders, & Emilie Vestergaard, Large Language Models, Small Labor Market Effects, NBER Working Paper No. 33777 (Sept. 2025), https://www.nber.org/papers/w33777.

Links survey data on ChatGPT adoption to administrative earnings and employment records in Denmark across 11 AI-exposed occupations. Finds essentially zero aggregate effects on earnings or hours worked through 2024 despite widespread, worker-reported adoption. Leverages Denmark’s comprehensive administrative registers to track individual-level labor outcomes with high precision. The null results contrast with large productivity gains observed in controlled experiments, aligning with a productivity J-curve dynamic in which organizational reorganization and task restructuring delay measurable economic effects. Shows that task-level automation does not translate mechanically into labor-market disruption. Helps explain the gap between micro-level productivity findings and macro-level employment outcomes.

• Jabarian, Brian, & Luca Henkel, Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews, Working Paper (2025), https://ssrn.com/abstract=5395709.

Reports results from a field experiment that replaced human recruiters with an AI voice agent for initial job interviews at a large recruiting firm. Finds that AI-led interviews increased job offers by 12%, job-start rates by roughly 18%, and 30-day retention by about 18% relative to human-conducted interviews. Shows that applicants offered a choice between AI and human interviewers predominantly selected the AI option. Attributes performance gains to reduced screening bottlenecks and improved candidate-position matching. Demonstrates that AI deployment can enhance labor-market efficiency and employment outcomes beyond cost reduction. Counters claims that AI adoption uniformly degrades hiring processes or worker experience.

• Johnston, Andrew C., & Christos A. Makridis, The Labor Market Effects of Generative AI: A Difference-in-Differences Analysis of AI Exposure, Working Paper (Oct. 25, 2025), https://ssrn.com/abstract=5375017.

Conducts a difference-in-differences analysis across sectors with varying levels of AI exposure to estimate labor-market effects. Uses the workplace rollout of AI tools beginning in 2021, combined with cross-sector variation in occupational exposure, to identify causal impacts. Finds that higher-exposure sectors recorded significant wage and employment gains, particularly among younger and more-educated workers. Interprets these patterns as evidence of labor-AI complementarity. Also identifies employment declines in sectors where AI directly substitutes for human labor. Highlights the heterogeneity of AI’s workforce effects, emphasizing that outcomes depend on whether the technology complements or replaces human inputs.

• Manne, Geoffrey A., Dirk Auer, Kristian Stout, Lazar Radic & Mario A. Zuñiga, Comments of the International Center for Law & Economics: DOJ Invitation to Comment on Promoting Competition in Artificial Intelligence, Int’l Ctr. for L. & Econ. (July 15, 2024), https://laweconcenter.org/wp-content/uploads/2024/07/Comments-of-the-International-Center-for-Law-DOJ-AI-RFI.pdf.

Presents a comprehensive critique of mandatory data-sharing and access mandates for AI models. Argues that incumbent data advantages do not constitute durable exclusionary barriers because model performance improves with data scale only up to a threshold, after which algorithmic innovation drives marginal gains. Warns that restrictions on partnerships between large technology firms and AI startups could reduce competition by limiting startup access to compute resources, technical infrastructure, and distribution channels. Frames such collaborations as mechanisms that facilitate entry and scaling, rather than entrench incumbent dominance.

• Merali, Ali, Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Translation, arXiv Preprint arXiv:2409.02391 (Dec. 10, 2024), https://arxiv.org/pdf/2409.02391.

Presents experimental evidence from a randomized controlled trial involving 300 professional translators to measure the economic effects of large language model adoption across varying training-compute levels. Establishes quantifiable “scaling laws” linking model capability to productivity outcomes. Finds that a tenfold increase in training compute reduced task-completion time by 12.3% and improved translation quality by 0.18 standard deviations. These gains produced a 16.1% increase in earnings per minute for participants. Identifies substantial heterogeneity in outcomes. Lower-skilled translators realized productivity gains roughly four times larger than those of higher-skilled peers, indicating that continued model scaling could raise aggregate productivity while narrowing skill-based wage differentials.

• Noy, Shakked, & Whitney Zhang, Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence, 381 Science 187 (2023).

Reports results from a randomized experiment involving 453 professionals performing writing tasks. The treatment group used ChatGPT-3.5, while the control group completed assignments without AI assistance. Finds a 40% reduction in task-completion time and an 18% improvement in output quality, as rated by independent evaluators. Quality gains concentrated among workers in the bottom half of the initial skill distribution, producing a measurable reduction in performance inequality. Participants also reported higher job satisfaction and self-efficacy when using AI tools. Provides controlled evidence that generative AI can materially increase knowledge-worker productivity while improving, rather than degrading, job quality. Demonstrates a skill-compression effect in which AI enables lower-skilled workers to approach expert-level performance.

• Organisation for Econ. Co-operation & Dev. (OECD), Artificial Intelligence and Competitive Dynamics in Downstream Markets, OECD Roundtables on Competition Pol’y Papers No. 331 (2025), https://www.oecd.org/en/publications/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_ccf0624a-en.html.

Analyzes AI’s effects on entry barriers and minimum-efficient scale across industries. Finds that AI enables smaller firms to automate functions—such as marketing analytics, customer service, and data processing—that previously required large, specialized teams. Examines vertical integration between model providers and application developers, identifying mixed competitive effects. Integration can foreclose rivals in some contexts, while improving coordination efficiency in others. Provides an international comparative perspective on AI-driven market structure. Clarifies how AI reshapes business organization and competitive dynamics beyond the foundation-model layer.

• Organisation for Econ. Co-operation & Dev. (OECD), How Are SMEs Using Generative AI? (2025), https://www.oecd.org/en/publications/generative-ai-and-the-sme-workforce_2d08b99d-en/full-report.html.

Presents findings from a 2024 cross-country survey of more than 5,000 small- and medium-sized enterprises examining generative AI adoption and workforce effects. Reports that roughly 31% of SMEs across the seven surveyed countries have adopted generative AI, with users citing improved employee performance and reduced workloads. Finds limited evidence of automation-driven job loss: 83% of adopting firms reported no change in overall staffing levels, instead using AI to address labor shortages and offset skill gaps. Identifies persistent barriers to broader diffusion. Nonadopters most frequently cited lack of business-use suitability (57%), legal and data-privacy concerns (54%), and insufficient internal workforce skills (50%). Suggests that while generative AI delivers productivity gains and can level competitive conditions, effective deployment currently favors firms with higher-skilled labor, potentially widening the SME digital divide.

• Peng, Sida, Eirini Kalliamvakou, Peter Cihon & Mert Demirer, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, arXiv Preprint arXiv:2302.06590 (2023), https://arxiv.org/pdf/2302.06590.

Reports results from a controlled trial evaluating GitHub’s Copilot AI coding assistant. Finds that developers using the tool completed programming tasks 55.8% faster than control-group participants. Identifies heterogeneous effects across user groups. Less-experienced developers, older programmers, and those who code more hours per day realized the largest productivity gains. Demonstrates substantial efficiency improvements in software development and highlights AI’s potential to lower skill barriers and support workforce entry into programming occupations.

• Schrepel, Thibault, & Jason Potts, Measuring the Openness of AI Foundation Models: Competition and Policy Implications, 34 & Comm. Tech. L. 279 (2025).

Evaluates the openness of 11 prominent AI foundation models—including GPT-4, Llama 3, and Gemini—by analyzing licensing structures through an innovation-commons framework, rather than purely technical criteria. Constructs an index of 18 variables grouped into knowledge sharing, anti-opportunism, and governance dimensions. Finds that openness operates along a spectrum, rather than a binary “open” versus “closed” classification. Most models cluster in the middle-to-lower range, with limited differentiation between proprietary systems and those marketed as open source. For example, Meta’s Llama 3 and OpenAI’s GPT-4 differ by only two index points. Shows that higher openness scores correlate with stronger upstream access to code and datasets. At the same time, most models lack robust participatory-governance mechanisms and safeguards against opportunism, regardless of whether they originate from large technology firms or smaller developers.

• Schwarcz, Daniel, Sam Manning, Patrick Barry, David R. Cleveland, J. J. Prescott & Beverly Rich, AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice, L. & Empirical Analysis (forthcoming 2026) (Dec. 6, 2025), https://ssrn.com/abstract=5162111.

Reports results from a randomized controlled trial assessing how advanced AI tools affect the quality and efficiency of legal work produced by law students. Compares two systems—a specialized legal platform using retrieval-augmented generation (Vincent AI) and a general-purpose reasoning model (OpenAI’s o1-preview)—against a no-AI control group. Finds that both tools significantly improved legal-document quality across most tasks, with gains concentrated in clarity, organization, and professional tone. The reasoning model produced the largest improvements in analytical depth, while the retrieval-augmented system reduced hallucinations and fabricated citations. Documents substantial productivity gains. Students completed assignments 50% to 130% faster with AI assistance, indicating that newer reasoning and retrieval architectures mitigate quality constraints observed in earlier-generation models.

• Siddik, Abu Bakkar, Yong Li & Anna Min Du, Unlocking Funding Success for Generative AI Startups: The Crucial Role of Investor Influence, 69 Rsch. Letters 106203 (2024).

Examines the determinants of funding success across 556 generative AI startups operating between 2010 and July 2024. Uses principal-component analysis and regression modeling to evaluate how financial and technological factors shape capital formation. Finds that investor influence—captured through variables such as investor count, lead-investor participation, and funding-round depth—exerts a significant positive effect on total funding raised. In contrast, technological influence, including IT spending and patent activity, shows no statistically significant relationship with funding outcomes. Concludes that investor networks and financing structures play a more decisive role in capital accumulation for generative AI ventures than technology metrics alone.

• Stout, Kristian, Comments of the International Center for Law & Economics: Request for Information on the Development of an Artificial Intelligence (AI) Action Plan, Int’l Ctr. for L. & Econ. (Mar. 14, 2025), https://laweconcenter.org/wp-content/uploads/2025/03/OSTP-AI-2025-comments-v-1.pdf.

Advocates a regulatory framework that balances risk management with the imperative to sustain U.S. innovation leadership. Warns that burdensome or premature rules could constrain the AI sector’s dynamic growth trajectory. Critiques international regulatory models, including the EU’s AI Act, and highlights the practical difficulty of defining artificial intelligence for legal purposes. Argues that overly broad or fragmented definitions risk undermining economic competitiveness and regulatory coherence. Calls for an evidence-based policy approach that accounts for open-source development dynamics and copyright-law considerations. Frames this calibrated governance model as necessary to preserve both technological safety and competitive vitality.

• Stout, Kristian, Federal Preemption and AI Regulation: A Law and Economics Case for Strategic Forbearance, WLF Legal Pulse (May 30, 2025), https://www.wlf.org/wp-content/uploads/2025/05/053025Stout_LP.pdf.

Argues for federal preemption of conflicting state-level AI regulations to prevent market fragmentation and excessive compliance costs. Examines Commerce Clause constraints on state laws that produce extraterritorial effects. Contends that a unified federal framework enabling interstate AI-service deployment would support broader adoption than a patchwork of incompatible state requirements. Applies a law & economics framework to questions of regulatory federalism. Clarifies legal constraints on state AI governance and articulates the economic case for federal harmonization.

• Struta, Iuri, GenAI Funding on Track to Set New Record in 2024, S&P Glob. Mkt. Intel. (Nov. 4, 2024), https://www.spglobal.com/market-intelligence/en/news-insights/research/genai-funding-on-track-to-set-new-record-in-2024.

Quantifies venture-capital investment trends in generative AI, reporting more than $20 billion invested in startups through Q3 2024, surpassing the pace recorded in 2023. Highlights record-scale fundraising by frontier-model firms, including funding rounds exceeding $8 billion for companies such as Anthropic as they compete with large technology incumbents. Finds that a substantial share of startups concentrate on application-layer and vertical-specific solutions, rather than foundation-model development. Provides market-structure evidence showing capital dispersion across the AI stack. Reflects strong investor optimism despite high capital-burn rates and uncertain profitability timelines among many entrants. Offers a financing and capital-allocation lens for assessing competitive dynamics in the AI sector.

• Teeselink, Bouke Klein, Generative AI and Labor Market Outcomes: Evidence from the United Kingdom, King’s Coll. London (Dec. 21, 2025), https://ssrn.com/abstract=5516798.

Uses a difference-in-differences framework centered on ChatGPT’s release to analyze U.K. labor-market outcomes. Finds that firms with higher large-language-model exposure reduced overall employment, with losses concentrated among junior and entry-level roles. Documents hiring contractions in technical and creative occupations, including software engineering and design, while demand for interpersonal roles—such as sales—remained stable or increased. Indicates that AI adoption reshaped workforce composition, rather than uniformly reducing labor demand. Shows that average firm-level compensation rose, driven by the shedding of lower-paid junior staff. At the same time, advertised salaries for exposed occupations declined, signaling weaker market demand for those skills. Finds that displacement effects concentrate in higher-wage labor segments, diverging from earlier automation waves that primarily affected low- to middle-skill routine work.

• Weinstock, Lida R., & Paul Tierno, The Macroeconomic Effects of Artificial Intelligence, CRS In Focus IF12762, Rsch. Serv. (Apr. 1, 2025), https://www.congress.gov/crs_external_products/IF/PDF/IF12762/IF12762.4.pdf.

Surveys estimates of AI’s macroeconomic effects, finding that projected GDP impacts vary widely but remain “typically positive.” Cites benchmark projections, including a Goldman Sachs estimate of a 0.9% cumulative GDP increase over 10 years and another study projecting long-run output gains of up to 35% above baseline. Emphasizes that realized economic effects depend on adoption rates and diffusion dynamics, noting that widespread integration may unfold over “multiple decades,” mirroring the trajectory of personal-computer deployment. Addresses labor-market implications, concluding that current evidence shows AI replacing tasks rather than entire jobs. Frames these dynamics as likely to influence productivity growth and income distribution over time.

• Yusuf, Shahid, The Macroeconomic Consequences of AI: The Next Great Divergence, N. Dev. Programme (2025), https://www.undp.org/sites/g/files/zskgke326/files/2025-12/the-macroeconomic-consequences-of-ai.pdf.

Investigates AI’s macroeconomic implications, focusing on its capacity to counter declining productivity and reshape income distribution across the Asia-Pacific region. Documents a sustained global slowdown in economic growth and total factor productivity since the early 2000s, noting that many high-growth Asian economies have relied more on capital accumulation than on efficiency gains. Frames AI as a potential response to this “productivity drought” while emphasizing uneven and time-lagged benefits. Finds that adoption effects will likely vary across sectors and labor groups, with gains materializing gradually rather than immediately. Warns that AI diffusion could widen within-country income disparities, reversing the decline in global inequality observed from the late 1990s through 2015. Positions distributional divergence as a central policy risk accompanying AI-driven productivity growth.

[1] When a general-purpose technology—such as electricity, computers, or artificial intelligence—emerges, firms must invest in intangible assets, including process redesign, worker training, and new software and business models. These investments absorb time and capital without generating immediate, measurable output, often producing an initial slowdown in recorded productivity. As complementary investments mature and integration deepens, productivity growth accelerates, forming the upward stem of the J-curve.

[2] Hulten’s Theorem holds that the aggregate economic effect of a productivity change in a given industry equals that industry’s share of national GDP, measured by total sales. The result implies that an activity’s economywide importance depends on its relative size, not on the complexity of its upstream or downstream linkages.

[3] A computable general-equilibrium (CGE) model simulates how shocks—such as new technologies, taxes, or trade policies—propagate across an economy. Using observed economic data, the model estimates how producers and consumers adjust until the system converges to a new equilibrium.

California lawmakers are considering a major change to the state’s antitrust laws that sounds technical but could have a simple and troubling result — higher prices for Californians.

Read the full piece here.

I. Introduction

We appreciate the opportunity to comment on the Competition & Markets Authority’s (CMA) call for evidence (the Call for Evidence)[1] in connection with its Merger Efficiencies Review. The International Center for Law & Economics (ICLE) is a nonprofit, nonpartisan global research and policy centre dedicated to building the intellectual foundations for sensible, economically grounded policy. ICLE promotes law & economics methodologies to inform public-policy debates and has longstanding expertise in competition law and policy. These comments aim to help ensure that competition law remains grounded in clear rules, established precedent, sound evidence, and economic analysis. ICLE has also submitted views on related issues to the Canadian Competition Bureau regarding its draft merger guidelines.[2]

We commend the CMA’s adoption of the ‘4Ps’ framework.[3] A merger-control regime focused on pace, predictability, proportionality, and process aligns with a law & economics approach and supports markets in delivering lower prices, higher quality, innovation, and consumer welfare.

The CMA’s review presents an opportunity to clarify how efficiencies should be assessed in practice. Because merger review is inherently prospective, both harms and benefits depend on projections about future competitive conditions. The framework should therefore treat uncertainty consistently, avoid systematically discounting longer-term benefits, and recognise that innovation, investment, and integration often generate consumer welfare over extended horizons.

Our recommendations are as follows:

• Preserve flexibility and analytical symmetry in the assessment of efficiencies, particularly with respect to timing. Where theories of harm rely on medium- or long-term projections, efficiencies over comparable horizons should be admissible and evaluated under similar evidentiary standards.
• Adopt a balanced approach to sufficiency by considering both likelihood and magnitude. Dynamic and innovation-related efficiencies should not be discounted relative to more immediate static effects.
• Interpret merger specificity pragmatically, focusing on whether efficiencies could realistically be achieved through less anticompetitive alternatives rather than requiring narrow ‘buyer-specific’ proof. Evidence from comparable transactions and industry experience should be relevant.
• Clarify that efficiencies are not suspect merely because they strengthen competitive performance. Efficiencies that allow firms to compete more effectively—through lower costs, improved quality, or greater innovation—benefit consumers even if rivals lose sales.
• Apply consistent and proportionate evidentiary standards to harms and efficiencies. Internal documents, forward-looking plans, and qualitative evidence should be probative when assessing efficiency claims, just as they are when assessing theories of harm.
• Give appropriate weight to dynamic efficiencies and innovation, particularly in sectors characterised by high fixed costs, rapid technological change, and complementarities. Context-specific analysis reduces the risk of deterring pro-investment and pro-innovation transactions.

The sections that follow elaborate on these points by addressing timeliness, the balancing of harms and benefits across time and customer groups, merger specificity, pass-through to consumers, evidentiary standards, and the treatment of dynamic efficiencies.

II. Framework for Assessing Merger Efficiencies

The CMA applies a structured framework to merger efficiencies. Once it identifies a realistic prospect of competitive harm, the merging parties must show that the claimed efficiencies are rivalry-enhancing, timely, likely and sufficient, and merger-specific. The resulting benefits must also accrue to UK consumers. This section addresses each of these analytical steps and offers recommendations to improve their practical application.

Our comments focus on how the framework operates in practice. First, the assessment should treat uncertainty symmetrically: the evidentiary standards applied to long-horizon harms should also apply to long-horizon efficiencies. Second, the CMA should adopt a transparent approach to weighing harms and benefits across time horizons, competitive parameters, and affected customer groups, including cases involving innovation and two-sided markets.

Third, merger specificity should be evaluated with attention to real-world organisational constraints. Contractual alternatives such as licensing or joint ventures may not replicate the integration benefits of a merger, and the analysis should not impose unduly narrow buyer-specificity requirements. Fourth, evaluation of consumer benefit should centre on residual competitive pressure, including entry conditions and potential competition, which determine whether efficiencies are passed through to consumers.

Finally, guidance should clarify that efficiencies are not suspect merely because they strengthen the merged firm’s ability to compete. Competition law protects the competitive process, not individual competitors. Clearer treatment of these issues would improve predictability, align the framework with legal and economic evidence, and encourage parties to present well-supported efficiency claims.

A. Temporal Symmetry in Assessing Efficiencies

The CMA seeks views on the factors it should consider when assessing the timeliness of merger efficiencies.[4] The current Merger Assessment Guidelines (MAGs) state that the CMA ‘will assess whether the claimed efficiencies are to be realised (and the resultant rivalry-enhancing effects felt) within the same timeframe as the CMA has adopted in the rest of its analysis’.[5] This language is constructive because it recognises a principle of temporal symmetry between harms and efficiencies. The MAGs do not prescribe a fixed timeframe, and that flexibility should be preserved.

The MAGs then qualify the symmetry principle: ‘usually the longer the time period necessary for efficiencies to be realised, the greater will be the level of doubt that efficiencies will be realised at all’.[6] This qualification effectively applies a rising scepticism discount to efficiencies. The Guidelines include no comparable qualification for theories of harm. Instead, they state that the CMA ‘will generally take a forward-looking approach to the assessment of any theories of harm, considering the effects of the merger both now, and in the future’.[7]

In practice, the CMA routinely evaluates long-horizon harms. When analysing potential competition—e.g., whether a firm ‘would have entered or expanded and could be expected to become a strong competitor’ absent the merger[8]—or assessing market-tipping risks in platform markets, the CMA projects competitive effects years ahead without applying an escalating doubt discount. The Guidelines therefore embed a structural asymmetry: long-term harms receive forward-looking treatment, while long-term efficiencies face progressive discounting based solely on time. The review should consider whether this one-directional scepticism is justified, or whether uncertainty should be treated symmetrically on both sides of the competitive assessment.

A more coherent approach—already implicit in the MAGs—is to apply equal evidentiary treatment. If a theory of harm relies on projected anticompetitive effects several years in the future, efficiencies expected over a comparable horizon should be admissible. Aligning the treatment of harms and benefits would improve analytical consistency and the credibility of the assessment.[9]

Timeliness should also reflect industry dynamics. In fast-moving sectors, market forces may mitigate short-term concerns, while longer-term competitive developments become decisive. Efficiencies may therefore materialise quickly and alleviate competitive risks. In industries with long development cycles, by contrast, a longer horizon for efficiencies is realistic and often necessary. As the OECD recognises, ‘mergers have a positive long-run impact on performance by creating efficiencies or synergies, even though such effects can take several years to materialize’.[10]

B. Weighing Harms and Efficiencies Across Time, Parameters, and Customers

The CMA also asks how it should evaluate claimed efficiencies when determining whether they offset potential anticompetitive effects, particularly where harms and benefits arise over different time horizons, affect different competitive parameters, or fall on different groups of customers.[11]

Balancing harms and efficiencies is inherently uncertain in merger enforcement. The task becomes more difficult when near-term harms must be weighed against more distant benefits. Static efficiencies may arise quickly but often generate modest welfare gains. Dynamic efficiencies—often associated with innovation—typically take longer to materialise but may produce substantially larger benefits.

One useful approach is analogous to risk-regulation frameworks that evaluate trade-offs along two dimensions: likelihood and impact. Short-term harms may be relatively certain yet limited in magnitude, while long-term efficiencies may be less certain but potentially transformative.[12] Explicitly assessing both probability and expected impact, to the extent feasible in a given case, would improve transparency and avoid systematically discounting longer-term efficiencies solely because of timing. It would also allow the CMA to capture the full value of efficiencies, while maintaining appropriate caution regarding short-term consumer harm.

Similar issues arise when harms and efficiencies affect different competitive parameters. A common example involves short-term price effects alongside longer-term innovation benefits. Such cases inevitably require policy judgement. Recent developments in UK competition policy—including greater emphasis on growth, investment, and innovation—reflect increasing recognition of dynamic competition. Careful evaluation of innovation-related efficiencies is therefore consistent with both economic evidence and evolving policy priorities.[13] The CMA’s assessment in the Vodafone/Three merger,[14] where potential price effects were considered alongside longer-term infrastructure investment and innovation commitments, illustrates the relevance of this balancing exercise.

Mergers may also affect customer groups unevenly, whether geographically or temporally. A merger may raise prices in certain local markets while enabling nationwide investment that benefits consumers more broadly. In other cases, environmentally oriented innovation may increase short-term costs while generating longer-term consumer benefits through improved products or reduced environmental harm.[15] Similar issues arise in two-sided markets, where disadvantages to one side of market participants may coincide with benefits to another.[16] These ‘out-of-market efficiency’ scenarios underscore the need to assess efficiencies across relevant customer groups and time horizons, rather than focusing narrowly on immediate effects.[17] Where such mismatches arise, the CMA’s policy mandate supports giving appropriate weight to nationwide and longer-term efficiencies.

C. Assessing Merger Specificity

The CMA also seeks views on how merger specificity should be assessed. The requirement reflects the principle that efficiencies must arise from the merger itself, rather than from less restrictive alternatives, such as licensing agreements or joint ventures.

Economic and management research suggests this assessment should be applied cautiously.[18] As the Call for Evidence recognises, practical barriers may prevent firms from achieving comparable efficiencies through contractual arrangements.[19] Transaction-cost economics shows that licensing agreements and joint ventures can be costly to negotiate, implement, and monitor.[20] They may also create dependence on jointly developed assets or specialised knowledge, increasing the risk of disputes or opportunistic behaviour. Anticipating these risks, firms may forgo such arrangements altogether, meaning the efficiencies would not materialise absent the merger.

Management literature likewise treats partnerships and mergers as distinct organisational modes, not functional substitutes. Partnerships typically involve complementary specialisation and limited coordination. Mergers enable deeper integration and broader deployment of resources.[21] While both structures can generate efficiencies, they are not reliably interchangeable.[22] The CMA should therefore avoid assuming that efficiencies achievable through integration can readily be replicated through contractual alternatives.

Efficiency claims should also not be required to be ‘buyer-specific’ in an unduly narrow sense. Where multiple plausible acquirers could generate comparable efficiencies, that should satisfy the merger-specificity requirement. The parties should also be able to rely on evidence from comparable transactions within the industry.

Finally, willingness to pay can be informative. If an acquirer offers a materially higher price than other plausible bidders, that may provide corroborating—though not conclusive—evidence that the buyer expects to realise greater value from the transaction. Such expectations often reflect complementarities, scale economies, or integration efficiencies that the acquirer believes it can deploy more effectively than alternative purchasers.

D. Pass-Through and Residual Competitive Pressure

Finally, the CMA seeks views on how it should assess whether efficiencies are likely to benefit consumers, particularly through the pass-through of cost savings or quality improvements.[23] This assessment is necessarily contextual.

The key factor is the degree of residual competitive pressure after the merger. Consumer harm arises when competitive pressure weakens; consumer benefit arises only where sufficient pressure remains to induce firms to pass on efficiency gains. Evaluating that pressure requires a case-specific assessment of market structure, entry conditions, and the transaction’s strategic rationale. Where a merger aims to strengthen competition against a larger incumbent or enable entry into new markets, the merged firm will often have strong incentives to translate efficiencies into lower prices, improved quality, or greater innovation. Where a merger substantially increases market power, pass-through becomes less likely.[24]

Entry conditions and potential competition therefore warrant particular attention. In technologically dynamic markets, competitive constraints often come not only from existing rivals but also from potential entrants and rapid innovation cycles. These forces can both discipline firms to pass on efficiencies and help explain why efficiency-seeking mergers occur in the first place.[25] Recognising these features would allow the CMA to evaluate efficiency claims more realistically in sectors characterised by rapid technological change.

E. Efficiencies, Rival Harm, and the Competitive Process

The Call for Evidence notes that some stakeholders have expressed concern that ‘the CMA might consider that the efficiencies could reduce rivals’ ability to compete effectively with the merged entity’, creating a ‘barrier to submitting efficiency claims’.[26]

There is no indication that this concern reflects actual CMA policy. Nonetheless, updated guidance could usefully clarify that efficiencies will not be treated negatively merely because they strengthen the merged firm’s competitive position. Historically, some authorities in the United States and the European Union were sceptical of merger-efficiency arguments and occasionally treated them as suspect.[27] That approach changed as agencies and courts adopted a more economic framework. As the U.S. Department of Justice explained: ‘challenging a merger because it will create a more efficient firm through economies of scale and scope is at odds with the fundamental objectives of the antitrust laws’.[28]

The relevant distinction is between efficiencies that disadvantage rivals by benefiting consumers—through lower prices, improved quality, or greater innovation—and conduct that harms rivals by distorting or blocking competition. The former is procompetitive and should be credited. Efficiencies that enable a firm to offer consumers a better deal reflect the competitive process operating as intended. As the DOJ further observed: ‘efficiency and aggressive competition benefit consumers, even if rivals that fail to offer an equally “good deal” suffer loss of sales or market share’.[29]

Discouraging such transactions would protect competitors rather than competition, which would be inconsistent with modern economic understanding and with the CMA’s stated objective of supporting growth and investment.

III. Evidentiary Standards for Efficiencies

Merger control differs from conduct investigations because it evaluates effects that have not yet occurred. The prospective nature of the analysis makes both the production and assessment of evidence more difficult. The CMA therefore asks what types and extent of evidence it should consider when assessing efficiencies, and whether evidential standards should vary depending on the nature of the claimed efficiencies.[30]

No single answer fits every case. Each merger arises in a distinct competitive setting, and the evidence needed to substantiate efficiencies will vary across transactions. The CMA should therefore retain a flexible approach to evidential submissions. Analytical symmetry provides a useful organising principle. Evidence the CMA treats as sufficiently probative when evaluating potential harms—such as internal business documents, forward-looking projections, and strategic planning materials—should also be admissible to substantiate efficiency claims. If internal documents support a theory of harm, particularly in innovation-focused cases, comparable materials should also support an efficiency claim.

The same principle applies to longer-term efficiencies. Such claims inevitably involve uncertainty, but uncertainty is not unique to efficiencies. The CMA routinely evaluates long-run theories of harm that depend on extended time horizons and uncertain outcomes. In those cases, the agency relies on coherent documentary evidence, industry analysis, and economic reasoning[31] Applying comparable evidential standards to efficiencies would improve consistency. Parties can mitigate uncertainty by grounding claims in contemporaneous business plans, sectoral analyses tracking expected technological developments, or investment commitments that credibly bind them to future action.[32] This evidence will not eliminate uncertainty, but it can demonstrate that claimed efficiencies are plausible and tied to observable commercial incentives.

The CMA also asks whether qualitative and quantitative evidence should receive different weight.[33] A general preference for either would be inappropriate. The relevant evidential mix depends on the nature of the claimed efficiency and the characteristics of the industry. Short-term cost savings often lend themselves to quantitative substantiation through cost modelling, projected scale economies, or financial forecasts. Some longer-term efficiencies may also be partly quantifiable, such as achieving minimum efficient scale or measurable productivity gains. Innovation-related efficiencies are less amenable to precise quantification. Claims that a merger will accelerate product development, enable new technological trajectories, or combine complementary capabilities will rely more heavily on qualitative evidence, including internal strategy documents and technological roadmaps.[34] Such evidence should not be discounted merely because precise quantification is difficult. Instead, the CMA should evaluate it in light of the transaction’s commercial rationale and the broader industry context.

A case-specific, context-sensitive approach, combined with symmetry between the treatment of harms and efficiencies, would improve both credibility and practical administrability. It would also reduce the risk that significant efficiencies are overlooked simply because they cannot be demonstrated with the same quantitative precision as short-term price effects. This approach aligns with the CMA’s objective of supporting growth and investment.

IV. Assessing Dynamic Efficiencies and Innovation Effects

The CMA’s final substantive question concerns the challenges of assessing and evidencing dynamic efficiencies. These are understood as benefits that generate ‘long-term benefit to customers by increasing the merged firms’ ability and incentive to innovate, invest or undertake research and development to improve the quality of products/services, introduce new products/services or improve production processes’.[35]

Mergers can affect competition not only through short-term price effects but also through innovation and investment. The CMA’s focus on dynamic efficiencies is therefore welcome. Assessing such efficiencies, however, raises conceptual and evidentiary issues distinct from those associated with static cost savings. The analytical framework should ensure that potentially significant innovation benefits are not systematically discounted.

Dynamic efficiencies are inherently uncertain. Predicting future benefits requires evaluating firms’ incentives, capabilities, and technological environment, rather than extrapolating from current market conditions. In this respect, dynamic-efficiency analysis resembles innovation-based theories of harm. Authorities regularly assess possible reductions in innovation rivalry using imperfect evidence. Comparable evidentiary tolerance is therefore appropriate when considering potential innovation benefits.

Mergers may enhance innovation where they combine complementary assets, capabilities, or resources that are difficult to coordinate contractually. Economic and management research emphasises that innovation often depends on integrating technological know-how, financial resources, data assets, and organisational capabilities within a single governance structure.[36] Where such complementarities exist, mergers can strengthen both the ability and incentives to invest in R&D. A merger that accelerates innovation or combines complementary capabilities may intensify competition over time, even if concentration increases in the short term. Gregory Sidak and David Teece similarly argue that a ‘neo-Schumpeterian’ approach places less weight on market share and greater weight on potential competition and firm capabilities.^[37]

Scale also matters in innovation-intensive sectors characterised by high fixed costs. Achieving sufficient scale may enable firms to undertake projects that would otherwise be technically or financially infeasible. Larger integrated firms may also better absorb the risks associated with uncertain innovation outcomes and therefore sustain long-term investment.[38]

Fixed-cost efficiencies may not immediately lower prices, but they affect investment incentives, risk tolerance, and capacity expansion. Focusing exclusively on short-term price effects risks understating long-run competitive benefits. Oliver Williamson demonstrated that merger analysis should weigh scale economies against deadweight loss to avoid significant economic waste.^[39]

Mergers can also facilitate access to intangible capital, including expertise, reputation, customer relationships, and technological infrastructure.[40] Smaller innovative firms often develop novel technologies but lack the resources required for commercialisation. Integration with a larger partner can accelerate deployment. This pattern appears frequently in technology-intensive industries, including pharmaceuticals, where acquisitions often support long-term development and market introduction.[41]

Evaluating dynamic efficiencies requires a broader evidentiary toolkit than static efficiency analysis. Quantification is often difficult because outcomes depend on uncertain technological trajectories, evolving consumer preferences, and rivals’ responses. Nonetheless, several forms of evidence can support such claims. Internal business documents and strategic plans—such as R&D roadmaps, investment plans, assessments of technological complementarities, and integration strategies—provide a natural starting point. Industry studies and sector reports can contextualise technological trends and competitive dynamics.[42] Historical evidence from comparable transactions may also offer useful benchmarks.

Merger specificity should also be assessed in a commercially realistic manner. The CMA asks whether efficiencies could be achieved ‘by other means’, including independent investment or licensing.[43] EU guidance similarly examines whether less anticompetitive but realistic alternatives would produce comparable efficiencies. In dynamic-efficiency cases, the central question often concerns timing and governance: whether alternative arrangements could plausibly deliver equivalent integration of complementary intangible assets and coordinated investment within the relevant timeframe.

Finally, qualitative evidence should not be discounted automatically. Innovation effects often manifest through changes in organisational capability, technological direction, or strategic positioning, which cannot readily be captured through numerical forecasts. A balanced evidentiary approach that considers both qualitative and quantitative sources will produce more accurate assessments.

V. Conclusion

Merger-efficiency assessment should rest on sound economic analysis and reflect the realities of modern markets. Evaluating efficiencies is inherently forward-looking and uncertain, particularly where investment, innovation, and long-run competitive dynamics are involved. Merger control should therefore avoid rigid or formalistic requirements that exclude relevant evidence or systematically discount longer-horizon benefits. A case-specific, context-sensitive approach provides a more reliable basis for decision-making.

Several principles follow. The CMA should apply analytical symmetry between harms and efficiencies: when theories of harm rely on projections about future entry, innovation, or market tipping, comparable projections should be acceptable when supported efficiency claims are presented. Efficiencies should not be discounted merely because they materialise over longer timeframes or resist precise quantification. Assessment should instead consider probability and expected magnitude, as well as residual competitive pressure that determines whether benefits are passed through to consumers.

Merger specificity also requires a commercially realistic application. The question is whether comparable efficiencies could plausibly be achieved through less anticompetitive alternatives, not whether they are unique to a particular buyer. Contractual arrangements, such as licensing or joint ventures, often cannot replicate the integration of complementary assets, governance, and investment incentives that a merger provides. Recognising this reduces the risk of false negatives that deter welfare-enhancing transactions. Guidance should also clarify that efficiencies are not suspect merely because they strengthen the merged firm’s ability to compete; competition law protects the competitive process, not individual competitors.

Evidentiary standards should remain flexible and consistent. Internal documents, business plans, industry analysis, and qualitative evidence may all be probative, particularly for dynamic efficiencies that cannot be reduced to short-term price effects. A balanced approach reduces the risk that innovation benefits, scale economies, and intangible-capital integration are overlooked.

These recommendations align with the CMA’s ‘4Ps’ framework. Clear and symmetric standards improve predictability for parties and stakeholders. Proportionate evidentiary expectations facilitate timely engagement on efficiencies and improve pace. A flexible, economically grounded analysis supports proportionality by calibrating enforcement to likely competitive effects. Finally, transparent treatment of efficiencies strengthens the review process and reinforces confidence that merger decisions promote growth, investment, and long-term consumer welfare.

[1] Competition & Markets Auth., Merger Efficiencies Review: Call for Evidence (15 January 2026), https://connect.cma.gov.uk/call-for-evidence-merger-efficiencies-review (‘Call for Evidence’).

[2] Ian Adams, Eric Fruits, Brian Albrecht, Daniel J. Gilman & Geoffrey A. Manne, Comments of the International Center for Law & Economics to the Competition Bureau of Canada: Proposed Merger Enforcement Guidelines (10 February 2026), https://laweconcenter.org/wp-content/uploads/2026/02/Competition-Bureau-Canada-Merger-Comments-2026.pdf.

[3] Call for Evidence, supra note 1, at 2.

[4] Call for Evidence, supra note 1, at 7.

[5] Competition & Markets Auth., Merger Assessment Guidelines 61 (18 March 2021), https://assets.publishing.service.gov.uk/media/61f952dd8fa8f5388690df76/MAGs_for_publication_2021_–_.pdf.

[6] Id. ¶ 8.12.

[7] Id. ¶ 2.10.

[8] Id. ¶ 2.18(b).

[9] See Mario Todino, Geoffroy van de Walle & Lucia Stoican, EU Merger Control and Harm to Innovation—A Long Walk to Freedom (from the Chains of Causation), 64 Antitrust Bull. 11 (2019) (making a similar call in the context of EU merger regulation).

[10] Org. for Econ. Co-operation & Dev., Merger Control in Dynamic Markets 12 (10 March 2020), https://www.oecd.org/en/publications/merger-control-in-dynamic-markets_d3752037-en.html.

[11] Call for Evidence, supra note 1, at 7.

[12] Dirk Auer, Innovation Defenses and Competition Laws: The Case for Market Power (2019) (Ph.D. dissertation, Univ. of Liège), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4667754.

[13] Ufuk Akcigit & John Van Reenen, The Economics of Creative Destruction: New Research on Themes from Aghion and Howitt (Harv. Univ. Press 2023).

[14] Competition & Markets Auth., Merger Investigation into the Anticipated Joint Venture Between Vodafone Group plc and CK Hutchison Holdings Ltd. Concerning Vodafone Ltd. and Hutchison 3G UK Ltd.: Case Closure Summary (2 June 2025), https://assets.publishing.service.gov.uk/media/683d54fffd325db61c5ff27d/case_closure_summary.pdf.

[15] Roman Inderst & Stefan Thomas, Prospective Welfare Analysis—Extending the Willingness-to-Pay Assessment to Embrace Sustainability, 18 J. Competition L. & Econ. 551 (2022).

[16] Francesco Ducci, Out-of-Market Efficiencies, Two-Sided Platforms, and Consumer Welfare: A Legal and Economic Analysis, 12 J. Competition L. & Econ. 591 (2016).

[17] The problem of ‘out-of-market’ efficiencies is not confined to the United Kingdom. See Geoffrey A. Manne & Kristian Stout, The Evolution of Antitrust Doctrine After Ohio v. Amex and the Apple v. Pepper Decision That Should Have Been, 98 Neb. L. Rev. 425 (2019) (discussing the issue from a U.S. perspective).

[18] See, e.g., Louis Kaplow, Improving Economic Analysis in Merger Guidelines, 39 J. Econ. Persps. 29 (2025); Howard Shelanski, Efficiency Claims and Antitrust Enforcement, in The Oxford Handbook of International Antitrust Economics (Roger D. Blair & D. Daniel Sokol eds., 2015).

[19] Call for Evidence, supra note 1, at 7.

[20] See, e.g., Michael Riordan & Oliver Williamson, Asset Specificity and Economic Organization, 3 Int’l J. Indus. Org. 365 (1985); Armen A. Alchian, Specificity, Specialization, and Coalitions, 140 J. Institutional & Theoretical Econ. 34 (1984).

[21] David Mowery, Joanne Oxley & Brian Silverman, Strategic Alliances and Interfirm Knowledge Transfer, 17 Strategic Mgmt. J. 77 (1996).

[22] John Hagedoorn & Bert Sadowski, The Transition from Strategic Technology Alliances to Mergers and Acquisitions: An Exploratory Study, 36 J. Mgmt. Stud. 87 (1999).

[23] Call for Evidence, supra note 1, at 7.

[24] Even here, a sufficiently radical innovation can induce a monopolist to pass benefits to consumers (e.g., by pricing below the pre-innovation competitive level). See Kenneth J. Arrow, Economic Welfare and the Allocation of Resources for Invention, in The Rate and Direction of Inventive Activity: Economic and Social Factors (Princeton Univ. Press 1962).

[25] Michael Gort & Steven Klepper, Time Paths in the Diffusion of Product Innovations, 92 Econ. J. 630, 646 (1982).

[26] Call for Evidence, supra note 1, at 6.

[27] William J. Kolasky & Andrew R. Dick, The Merger Guidelines and the Integration of Efficiencies into Antitrust Review of Horizontal Mergers, U.S. Dep’t of Justice (2003), https://www.justice.gov/sites/default/files/atr/legacy/2007/07/11/11254.pdf.

[28] U.S. Dep’t of Justice, Antitrust Div., Range Effects: The United States Perspective — Antitrust Division Submission for OECD Roundtable on Portfolio Effects in Conglomerate Mergers 3 (12 October 2001), https://www.justice.gov/sites/default/files/atr/legacy/2015/01/26/9550.pdf.

[29] Id. at 2.

[30] Call for Evidence, supra note 1, at 7–8.

[31] See Yossi Spiegel, The Facebook–Giphy Merger, Rev. Indus. Org. (forthcoming).

[32] The CMA adopted a similar approach in the recent Vodafone/Three merger. Comparable recommendations also appear in Mario Draghi’s report on European competitiveness.

[33] Call for Evidence, supra note 1, at 8.

[34] For example, in a prominent merger decision involving innovation considerations, the European Commission relied on extensive documentary evidence. See Nicholas Levy & Vassilena Karadakova, The EC’s Increasing Reliance on Internal Documents under the EU Merger Regulation: Issues and Implications, 39 Eur. Competition L. Rev. 12 (2018).

[35] Call for Evidence, supra note 1, at 8.

[36] David J. Teece, The Foundations of Enterprise Performance: Dynamic and Ordinary Capabilities in an (Economic) Theory of Firms, 82 Acad. Mgmt. Persps. 328 (2014).

[37] Gregory Sidak & David J. Teece, Dynamic Competition in Antitrust Law, 5 J. Competition L. & Econ. 581 (2009).

[38] See, e.g., Brian Albrecht, Geoffrey A. Manne, David J. Teece & Mario Zúñiga, From Moore’s Law to Market Rivalry: The Economic Forces That Shape the Semiconductor Manufacturing Industry, Int’l Ctr. for L. & Econ., at 30 (12 November 2025), https://laweconcenter.org/resources/from-moores-law-to-market-rivalry-the-economic-forces-that-shape-the-semiconductor-manufacturing-industry (‘… standard indicators of market power—high market shares, long-term contracts, and customer relationship “stickiness”—are likely misleading in the semiconductor context. These features often reflect the natural result of technological and economic forces rather than strategic behavior designed to exclude rivals. The semiconductor industry’s recurring pattern of competition for the market is characterized by fierce competition for technological leadership that resets with each generation, rather than a traditional winner-take-all race susceptible to permanent monopolization’).

[39] Oliver E. Williamson, Economies as an Antitrust Defense: The Welfare Tradeoffs, 58 Am. Econ. Rev. 18 (1968) (‘[I]n the occasional case where efficiency and market power consequences exist, can economies be dismissed on the grounds that market power effects invariably dominate? If they cannot, then a rational treatment of the merger question requires that an effort be made to [estimate the tradeoffs].’).

[40] Nicolas Crouzet, Janice Eberly, Andrea Eisfeldt & Dimitris Papanikolaou, The Economics of Intangible Capital, 36 J. Econ. Persps. 29 (2022).

[41] Melanie Büssgen & Tom Stargardt, To Merge or Not to Merge? The Impact of Mergers and Acquisitions on Corporate Success in the Pharmaceutical Industry, 45 Managerial & Decision Econ. 2196 (2024).

[42] Jorge Padilla, Douglas H. Ginsburg & Koren Wong-Ervin, Dynamic Competition and Antitrust: Quick-Look Inferences from the Analysis of Big Tech’s R&D Expenditure Ratios, 86 Antitrust L.J. 897 (2025).

[43] Call for Evidence, supra note 1, at 6.

The International Center for Law & Economics (ICLE) respectfully submits these comments regarding the proposed merger between Charter Communications and Cox Communications, now under review by the California Public Utilities Commission (CPUC). ICLE is a nonprofit, nonpartisan research center that applies law & economics methods to public policy. Our work seeks to ensure that competition policy and regulation rest on sound economic analysis and promote consumer welfare, particularly in dynamic, technology-driven markets such as media and telecommunications.

ICLE has written extensively about this transaction and would welcome the opportunity to provide the Commission our full issue brief.[1] As explained in that brief, the proposed merger promotes competition and consumer welfare under the Clayton Act and satisfies the Federal Communications Commission’s public-interest standard. It likewise advances the public interest in California, consistent with the Commission’s longstanding review under Cal. Pub. Util. Code § 854(a).[2]

At the national level, the transaction represents geographic expansion, not horizontal consolidation. The same holds in California. Charter and Cox largely serve distinct territories, while the combination offers potential consumer benefits from economies of scale and strengthened multimodal competition.

Although Charter and Cox have approximately 25,503 overlapping locations,[3] that figure represents an exceedingly small share of broadband service statewide. In four California markets alone, Charter passes more than 940,000 locations, underscoring that the overlap is de minimis relative to the statewide footprint.[4] These overlap estimates also typically rely on the availability of gigabit service,[5] which risks overstating competitive significance by assuming uniform consumer demand for the highest-speed tiers. In reality, broadband demand is heterogeneous, and many households meet their needs at lower speed levels.[6]

Even within the limited overlap areas, most consumers face substantial intermodal competition. Fiber providers and rapidly expanding fixed-wireless offerings increasingly constrain cable pricing and investment decisions,[7] and these technologies deploy quickly. In some locations, low-earth-orbit satellite services provide an additional alternative. Taken together, these market dynamics substantially limit any plausible concern that the transaction will meaningfully reduce competition in California broadband markets.

The merger is also expected to generate significant efficiencies that benefit California consumers. Broadband deployment and network upgrades require large capital expenditures and face regulatory, construction, and demand uncertainty. These risks increase the cost of capital and can delay or deter investment. By combining operations, the merged firm can internalize efficiencies, eliminate duplicative costs, and spread fixed investments across a larger customer base, lowering operating costs and investment risk.

Those efficiencies improve the business case for accelerated upgrades to newer cable standards, enhancing service quality for existing customers and supporting expanded deployment in areas that currently lack robust broadband options or face limited cable competition from fiber or fixed-wireless providers. For California specifically, the same efficiencies make network-hardening investments more likely in wildfire-prone regions, helping maintain connectivity during and after emergency events.

The benefits extend beyond fixed broadband. The merger would allow Charter to offer competitive mobile service in Cox service areas using Spectrum Mobile and Wi-Fi offload capabilities. The combined firm would provide an additional competitor to AT&T, Verizon, and T-Mobile. Cox video subscribers would also gain access to Charter’s video platform and features.

As detailed in the attached brief, the transaction serves the public interest in California. Ratepayers are unlikely to face higher prices, and the merger eliminates little, if any, direct competition. A stronger combined firm can more effectively deploy and upgrade high-speed internet infrastructure. Consumers also gain additional options for video and mobile services, which should intensify competition and place downward pressure on prices.

ICLE would welcome the opportunity to discuss the transaction’s consumer effects further. We attach our issue brief, “The Competitive Effects of the Proposed Charter/Cox Transaction,” for the Commission’s consideration and appreciate the opportunity to comment.

[1] Eric Fruits, Ben Sperry & Kristian Stout, The Competitive Effects of the Proposed Charter/Cox Transaction, Int’l Ctr. for L. & Econ. (Sept. 17, 2025), https://laweconcenter.org/resources/the-competitive-effects-of-the-proposed-charter-cox-transaction.

[2] Cal. Pub. Util. Code § 854(a); see also Joint Reply of Charter Commc’ns, Inc., Charter Commc’ns Holdings, LLC, Cox Enters., Inc. & Cox Cal. Telecom, LLC (U05684-C) to Protests at 3, Joint Application of Charter Commc’ns, Inc., Charter Commc’ns Holdings, LLC & Cox Enters., Inc. for Approval Pursuant to Public Utilities Code Section 854 of the Indirect Transfer of Control of Cox California Telecom, LLC (U-5684-C) (July 30, 2025).

[3] Jeff Baumgartner, FCC Urged to Slap Permanent Bans on Data Caps, Paid Peering on Charter-Cox Deal, LightReading (Nov. 21, 2025), https://www.lightreading.com/cable-technology/fcc-urged-to-slap-permanent-bans-on-data-caps-paid-peering-on-charter-cox-deal.

[4] Pub. Advocates Off., Cal. Pub. Utils. Comm’n, Technical Appendices to Broadband Competition and Pricing Strategies in California’s Urban Markets (Jan. 14, 2026), https://www.publicadvocates.cpuc.ca.gov/-/media/cal-advocates-website/files/press-room/reports-and-analyses/260114-public-advocates-appendices-to-broadband-competition-and-pricing-strategies-paper.pdf.

[5] Id.

[6] See Eric Fruits, Geoffrey A. Manne, Ben Sperry & Kristian Stout, Dynamic Competition in Broadband Markets: A 2024 Update Int’l Ctr. for L. & Econ. (June 4, 2024), https://laweconcenter.org/resources/dynamic-competition-in-broadband-markets-a-2024-update; Geoffrey A. Manne, Kristian Stout & Ben Sperry, A Dynamic Analysis of Broadband Competition: What Concentration Numbers Fail to Capture, Int’l Ctr. for L. & Econ. (June 3, 2021), https://laweconcenter.org/resources/a-dynamic-analysis-of-broadband-competition-what-concentration-numbers-fail-to-capture.

[7] Jeff Baumgartner, FWA Subscriber Share Expands Even in Areas with Multiple Fixed Wireless Options, LightReading (Oct. 22, 2025), https://www.lightreading.com/fixed-wireless-access/fwa-subscriber-share-expands-even-in-areas-with-multiple-fixed-wireless-options.

Executive Summary

The International Center for Law & Economics (ICLE) welcomes the opportunity to submit these comments to the Competition and Markets Authority (CMA) regarding the proposed Conduct Requirements (CRs) for Google’s general search services. ICLE is a nonprofit, nonpartisan research centre dedicated to promoting economically grounded public policy. Our work applies law & economics analysis to competition policy, including digital markets, artificial intelligence, and platform regulation. We seek to ensure that competition law rests on clear legal standards, established precedent, and robust empirical evidence, and that regulatory interventions protect consumers, rather than particular competitors.

The CMA’s consultation arises at a pivotal moment for the UK digital economy. Google’s designation with Strategic Market Status (SMS) triggers the Digital Markets, Competition and Consumers Act 2024’s ex ante regulatory framework. While the DMCC Act aims to promote competition and innovation, the specific measures proposed here risk producing the opposite effect, if not carefully calibrated. Across the consultations, the central issue is proportionality: whether the proposed interventions address demonstrable consumer harm or instead impose costs on product quality, innovation, and user experience in order to alter competitive relationships among firms.

Our analysis draws on two established competition-law principles. First, the error-cost framework recognises that regulatory decisions under uncertainty can produce harmful false positives, particularly in dynamic markets where innovation is cumulative and path-dependent. Second, competition policy protects competition, rather than competitors. Conduct that disadvantages rivals—such as integrated search features, direct answers, or ranking differentiation—may nonetheless benefit users by reducing search costs and improving relevance. These principles are especially important in digital and AI-enabled markets characterised by rapid technological change, dispersed knowledge, and evolving consumer preferences.

We then evaluate the three proposed CRs.

Publisher Conduct Requirement: The proposed controls on the use of publisher content in generative-AI services risk distorting competition by imposing obligations on a single firm, while other AI developers remain subject only to generally applicable law. The measure also seeks to resolve disputes about the permissible use of copyrighted material—questions more appropriately addressed through intellectual-property law. In addition, the opt-out mechanism may degrade AI-generated search responses and fragment the user experience, while less distortive tools such as transparency and attribution could address many of the CMA’s concerns.

User Choice Conduct Requirement: The proposal for recurring choice screens and complex switching journeys rests on behavioural assumptions that receive limited support from real-world evidence. Empirical studies of prior remedies, including the Microsoft browser ballot and the Android choice screen, show modest and short-lived effects on market outcomes. Repeated prompts may therefore impose recurring time and attention costs on millions of users who already have settled preferences, without materially improving contestability.

Fair Ranking Conduct Requirement: The proposed obligation of ‘objective’ and ‘non-discriminatory’ ranking misunderstands the nature of search, whose value depends on selective curation and integration. The CMA has not identified direct evidence that individual ranking decisions are unfair, yet the CR would impose complex compliance obligations that are difficult to administer in machine-learning systems. Restrictions on integrated features may also eliminate efficiencies that reduce search costs and improve relevance, potentially degrading product quality without clear competitive benefits.

Taken together, the proposals risk substituting regulatory design for market processes, addressing uncertain or indirect harms with measures that impose immediate and durable costs. We therefore encourage the CMA to adopt a proportionate and evidence-based approach, to rely on demonstrable consumer harm rather than changes affecting particular rivals, and to prefer less distortive measures—especially transparency and attribution—over restrictions on product design and innovation.

I. Analytical Framework for Assessing the Proposed Conduct Requirements

The Digital Markets, Competition and Consumers Act 2024 (DMCC Act) marks a structural reorientation of UK competition policy, shifting from primarily ex post enforcement toward an ex ante regulatory regime for firms designated with Strategic Market Status (SMS). Although Parliament has now settled the legislative framework, the CMA retains substantial discretion in designing and implementing Conduct Requirements (CRs). That discretion remains bounded by statutory obligations, including proportionality, evidence-based analysis, and a focus on consumer outcomes, rather than the commercial position of particular competitors.

These consultations therefore raise questions not only about specific remedies, but also about how enforcement should operate in dynamic and technologically complex markets. In such settings, regulatory decisions must be made under persistent uncertainty. The law & economics literature has long recognised that intervention entails a risk of error, and that enforcement policy should account for both the likelihood and the cost of mistaken intervention. Where innovation is cumulative and path-dependent—as in digital and AI-enabled services—the cost of prohibiting beneficial conduct may be durable, while some competitive harms may dissipate through entry and technological change. This asymmetry makes careful calibration of intervention especially important.

A related concern is analytical focus. Competition policy seeks to protect competition, not particular competitors. Conduct that disadvantages rivals—such as integrated features or direct answers in search—may nonetheless benefit users by improving relevance, reducing search costs, and accelerating access to information. Assessing the lawfulness of such practices therefore requires close attention to demonstrable consumer harm, rather than changes in traffic, revenues, or other outcomes affecting rival firms.

The CMA’s proposed CRs should be evaluated against these principles. The central question is whether the proposed measures address clear and substantiated consumer harm in a proportionate manner, or whether they risk restricting product design and innovation in circumstances characterised by technological change, dispersed knowledge, and evidentiary uncertainty.

A. The Error-Cost Framework and Enforcement in Digital Markets

Central to our analysis is the ‘error-cost’ framework, a foundational contribution of law & economics scholarship to the design of competition-law institutions. Developed by Judge Frank Easterbrook and extended by scholars including Geoffrey Manne and Joshua Wright, the framework recognises that regulatory decision-making under uncertainty inevitably produces errors. The appropriate enforcement posture therefore seeks to minimise the total expected cost of those errors.[1] In complex and dynamic markets, authorities must weigh both the probability and the consequences of two distinct categories of error.

In competition enforcement, Type I errors (false positives) arise when authorities condemn conduct that is pro-competitive or benign. The economic consequences of over-enforcement can be durable: prohibitions may eliminate a business practice, chill experimentation, and foreclose associated efficiencies. Type II errors (false negatives), by contrast, occur when authorities fail to sanction anticompetitive conduct, potentially leading to supracompetitive prices or reduced output. As Judge Easterbrook observed, such errors may be partly self-correcting, because the prospect of monopoly profits attracts entry and competitive responses that erode market power over time.[2]

In digital markets, there are strong reasons to expect Type I errors to be particularly costly.[3] As Manne notes, uncertainty is ‘further magnified when antitrust decisions are made in innovative, fast-moving, poorly understood, or novel market settings’.[4] Innovation in these markets is cumulative and path dependent. If a regulator mistakenly proscribes a business model (e.g., integrated product features sometimes described as self-preferencing) or restricts a consumer-valued innovation (e.g., AI-generated overviews in search) the resulting welfare loss may be difficult to reverse. The foreclosed line of development may not re-emerge once firms redirect investment elsewhere. In this way, Type I errors can durably distort innovation incentives in dynamic industries.

The compounding nature of innovation losses reinforces this asymmetry. As John Yun explains, even modest regulatory drag on the rate of technological progress can generate substantial long-term welfare losses.[5] Measures that merely slow innovation—without stopping it—still impose accumulating costs over time. Consumers bear these costs through foregone improvements in quality, functionality, and price.

By contrast, the effects of Type II errors may be moderated in markets characterised by rapid technological change. Supracompetitive returns attract entry and encourage disruptive innovation. The history of the digital economy—including the displacement of AOL by Google, Yahoo by Facebook, and MySpace by later social-media platforms—illustrates that apparently durable positions can erode when confronted by superior technology or business models.[6] This does not suggest that enforcement is unnecessary. Rather, it indicates that dynamic adjustment should inform the calibration of intervention, particularly where evidence of durable consumer harm remains uncertain.

The CMA’s current proposals, particularly those addressing AI-related features and search-ranking conduct, risk underweighting Type I error costs. Several proposed conduct requirements assume that the authority can reliably distinguish between ‘fair’ and ‘unfair’ product-design choices in a setting characterised by emergent algorithmic outputs and rapid technological change. Yet, as F.A. Hayek observed, regulators necessarily lack access to the dispersed and contextual knowledge embedded in market processes.[7] The error-cost framework therefore supports a cautious approach: intervention should focus on clear and demonstrable consumer harm, rather than speculative concerns about potential competitive disadvantage.

B. Competition, Not Competitors

A recurring theme in the CMA’s consultation documents is the attention afforded to ‘publishers’ and ‘rival search engines’—entities that often stand in a partly competitive and partly complementary relationship with the SMS firm. Promoting a contestable market is a legitimate regulatory objective. The analysis must nonetheless avoid drifting toward a ‘competitor-welfare’ approach, in which the practical aim becomes preserving rivals’ revenues, rather than improving consumer outcomes.[8] Competition law has long distinguished between protecting competition and protecting competitors. Even in Brown Shoe Co. v. United States, the U.S. Supreme Court stated that ‘[i]t is competition, not competitors, which the [Sherman] Act protects’.[9]

This distinction has concrete implications. As Manne and Wright explain in the search context, conduct that disadvantages competitors—such as presenting a direct answer or an integrated feature above a link to a third-party website—may benefit consumers by reducing search costs and improving relevance.[10] When a search engine displays a maps result, a knowledge panel, or an AI-generated summary, users obtain information immediately, rather than through additional navigation, delay, and cognitive effort. The resulting reduction in traffic to third-party sites reflects competition on the merits, not necessarily anticompetitive foreclosure.[11]

Restrictions designed to preserve traffic flows or advertising revenues for incumbent publishers would therefore risk prioritising competitor interests over consumer welfare. The empirical literature on self-preferencing remains mixed and context-specific, and vertical integration often produces measurable consumer benefits.[12] The CMA’s assessment should accordingly focus on demonstrable consumer harm, rather than the commercial impact on rival firms.

II. Assessment of the Proposed Publisher Conduct Requirement

The CMA’s proposed Publisher Conduct Requirement (Publisher CR) governs Google’s use of publisher content—crawled for general search—in generative-AI services such as AI Overviews, AI Mode, and the Gemini assistant.[13] The proposal has three components. First, opt-out controls would allow publishers to withhold ‘Search Content’ from grounding in search generative-AI features and from training and grounding in broader AI services. Second, transparency obligations would require disclosure of how content is used and provision of engagement metrics to publishers. Third, attribution requirements would require reasonable steps to identify and credit publisher material in AI-generated responses.[14]

The CMA’s theory of harm rests on a leveraging concern: because Google holds SMS in general search, publishers allegedly have ‘no realistic option but to allow their content to be crawled’, enabling Google to extend that content into AI applications without meaningful choice.[15] Even accepting bargaining asymmetry in search, the proposed CR raises three concerns.

First, it risks distorting competition in generative-AI services by imposing obligations on a single firm while other AI developers relying on similar publicly available content remain outside the regime. Second, it addresses disputes about the use of copyrighted material that are more appropriately resolved through generally applicable intellectual-property law, rather than firm-specific competition rules. Third, the design of the controls may fragment AI-generated search responses and reduce their usefulness to users, undermining features intended to lower search costs.

These issues bear directly on proportionality. Intervention should address demonstrable consumer harm, not redistribute value among market participants or indirectly resolve copyright disputes, and the CMA should consider whether less distortive measures—particularly transparency and attribution—could achieve its objectives without degrading product quality or altering competitive conditions.

A. Regulatory Asymmetry in Generative-AI Markets

The proposed Publisher CR would create a marked regulatory asymmetry in generative-AI services. Google would face legally binding obligations to provide granular opt-out controls for training and grounding, alongside transparency, reporting, and attribution requirements.[16] Competing providers of generative-AI systems—including OpenAI, Anthropic, Meta, and a growing ecosystem of open-source foundation-model developers—would face no equivalent constraints, because they are not designated with SMS in general search and therefore fall outside the DMCC regime.

The competitive implications are significant. Output quality in generative-AI systems depends on the breadth and recency of data used for training and grounding. If publishers exercise opt-out rights at scale—an outcome the low-friction design of the controls may facilitate—Google would either exclude large portions of web content from AI-generated search responses or produce responses with material gaps. Either outcome would reduce the quality of Google’s AI Overviews and AI Mode relative to rival services that may continue to rely on the same publicly available web content without restriction. The proposal would therefore not merely affect product design; it would alter competitive conditions between providers of generative-AI services.[17]

This does not imply that Google lacks competitive advantages in general search, including efficient access to crawled content. It does indicate that proprietary search-index data is not a prerequisite for building competitive generative-AI products. Leading frontier models—ChatGPT, Claude, Llama, and DeepSeek—were developed without access to a dominant search index. Open datasets such as Common Crawl, together with commercial licensing arrangements, supply training material across the industry.[18] As prior ICLE submissions to this Authority and the U.S. Department of Justice note, ‘being the firm with the most data appears to be far less important than having enough data’, a threshold accessible to many firms.[19] Consistent with this, generative-AI markets exhibit rapid entry and substantial investment, and enforcement authorities have not identified concrete anticompetitive harm arising from incumbents’ data holdings in AI markets.[20]

International experience also illustrates the risk of competitive distortion from asymmetric digital-market regulation. Under the EU Digital Markets Act, Google removed certain integrated features, including clickable map modules and embedded previews, from European search results to comply with self-preferencing rules. Reports indicate a slower and more fragmented user experience, without clear competitive gains for rivals.[21]

The proposed Publisher CR could produce similar effects. By introducing gaps in AI-generated search responses where opted-out content would otherwise appear, the measure risks degrading the UK search experience while leaving competing generative-AI providers unaffected.

B. Copyright Issues Are Not Competition Issues

The Publisher CR appears motivated in part by concern that Google uses content crawled for search in AI applications without adequate publisher consent. The underlying issue—whether the use of publicly available web content for AI training and grounding infringes copyright—is, however, a matter of intellectual-property law, rather than competitive conduct.

Whether an AI developer’s use of copyrighted material constitutes ‘fair dealing’ under UK law (or ‘fair use’ under U.S. law) applies uniformly across firms, irrespective of market position.[22] A start-up training a foundation model on publicly available web content confronts the same copyright question as Google. By using the DMCC framework to create what is effectively an opt-out regime applicable only to a designated firm, the CMA risks addressing an intellectual-property dispute through a competition-law instrument.

This approach would produce a two-tier structure of rights and obligations. Google would need to secure effective consent—through opt-out mechanisms—to use publisher content for AI grounding, while other AI developers would remain subject only to generally applicable copyright law. If grounding AI search responses in publicly available web content constitutes fair dealing, Google should be permitted to do so on the same basis as other market participants. If it does not, the appropriate remedy lies in copyright enforcement applied consistently across firms, rather than in an asymmetric competition-law obligation imposed on a single company.

C. Technical Design and Consumer Effects of the Opt-Out Controls

The proposed controls also raise concerns about technical implementation and user impact. The Publisher CR distinguishes among three uses—training foundation models, fine-tuning derivative models, and grounding AI-generated responses—and applies different obligations to each.

The CMA acknowledges Google’s submission that fine-tuning ‘helps the model learn how to process information rather than what current information to display’ and that permitting opt-outs from fine-tuning would ‘raise the risk of downranking or mis-ranking publisher content in organic search results’.[23] The Authority has therefore provisionally excluded fine-tuning of search AI models from the opt-out regime. This reflects a proportionate assessment of product functionality. Similar considerations, though operating differently, arise in relation to grounding.

AI Overviews and AI Mode rely on retrieval-augmented generation, in which responses are corroborated by reference to material retrieved from the search index at query time. Allowing publishers to opt individual pages out of grounding, while remaining indexed for general search, would effectively give them a page-by-page veto over AI-generated responses. The likely consumer effect is a fragmented experience: authoritative information may appear as traditional links but be absent from the AI-generated summary.[24] Because these features are intended to reduce search costs, such fragmentation would reduce their usefulness and quality. The CMA should weigh these foreseeable user costs against the more speculative benefits of the opt-out mechanism.[25]

For these reasons, the Publisher CR would benefit from revision. First, the controls should not extend to AI training. Questions about the use of publisher content for model training are properly addressed through generally applicable intellectual-property law rather than firm-specific competition regulation. Second, the CMA’s proportionality analysis should account for competitive asymmetry. The CR imposes binding obligations on Google alone, while rival AI developers using similar web content remain unconstrained, creating incentives for publishers to withhold content selectively. Third, where intervention is warranted, priority should be given to transparency and attribution. Clear disclosure of how content is used, and proper attribution in AI-generated responses, would address many of the CMA’s concerns, while avoiding the product degradation and competitive distortion associated with the proposed opt-out controls.

III. Assessment of the Proposed User Choice Conduct Requirement

The proposed User Choice CR—including periodic choice screens, information screens, and ‘test-drive’ functionalities—rests on the premise that such interventions reliably promote competition.[26] The CMA identifies two objectives: ‘(i) increasing competition in general search services and (ii) having more people use a general search provider that better matches their preferences’.[27]

The consultation states that the obligations are supported by evidence showing that choice screens increase engagement, improve comprehension, and enable active decision-making in digital markets.[28] This evidence relies substantially on experimental and survey-based studies, including research by Mozilla[29] and recommendations issued by Bureau Européen des Unions de Consommateurs (BEUC),[30] which examine user interaction with designed choice interfaces. By contrast, longitudinal analyses of observed market behaviour receive less weight. Evidence from earlier EU remedies, including the Microsoft browser ballot and the Android choice screen, indicates more limited effects on market shares. These studies generally find modest changes in usage patterns, suggesting that users often continue to select their preferred service when presented with alternatives.

As Omar Vásquez Duque observes in a recent empirical assessment:

[A] key assumption behind choice screens is that consumer inertia sustains market dominance. However, the findings here suggest that consumers may not be as inert as conventionally assumed…This raises questions about the “true” effectiveness of choice screens… If effectiveness is defined as “encouraging users to consider alternative options,” and the browsers’ assessments are accurate, then choice screens have been partially successful. However, if the goal is to increase market contestability, the evidence suggests that choice screens alone are ineffective.[31]

These mixed findings raise proportionality concerns. Requiring recurring prompts imposes certain costs on users, including time and interruption, while the benefits to competition remain uncertain. Behavioural research also indicates that repeated prompts may produce habituation rather than deliberation, and that users with settled preferences may treat such screens as obstacles rather than meaningful opportunities to choose.

The CMA acknowledges that ‘focusing solely on levels of switching to alternative providers may be a misleading measure of a choice screen’s effectiveness’.[32] That observation underscores the central question: if switching is not the relevant metric, it is unclear how the intervention’s benefits should be measured or whether the associated user costs are justified.

A. Empirical Evidence on the Effectiveness of Choice Screens

The CMA’s proposal assumes that presenting users with a choice will materially shift market shares. The empirical record from EU choice-screen remedies suggests otherwise: observed changes in share are typically small and rarely exceed a few percentage points.[33]

A relevant precedent is the Android choice screen introduced following the European Commission’s Google Android decision.[34] An econometric analysis by Francesco Decarolis, Muxin Li, and Filippo Paternollo, using difference-in-differences methods, finds a statistically significant reduction in Google’s mobile search share in the European Economic Area (EEA). The magnitude, however, is modest—less than one percentage point in the headline estimate, with some variation across specifications.[35] Although rivals with greater pre-remedy awareness (proxied by desktop share) benefited somewhat more, the overall effect falls short of materially deconcentrating the market.[36]

Market-share data point in the same direction. Figure 1 reports European search-engine shares from January 2015 to May 2025, based on Statista data derived from the StatCounter tracking environment.[37] Over this period, Google’s share remains relatively stable, generally between roughly 90 and 95 per cent. Competing providers, including Bing, Yandex, and Yahoo!, remain clustered at low levels and rarely exceed a combined 5 to 8 per cent share.

FIGURE 1: Search Engine Market Share in Europe, January 2015 to May 2025

SOURCE: Statista

The CMA has acknowledged similar outcomes in its SMS Decision:

Since August 2019, following the European Commission’s Google Android decision, Google has introduced choice screens for general search providers on all new Android phones in the EEA and UK. However, notwithstanding Google’s submissions that this created opportunities for third-party providers to be set as the default (see paragraph 5.164(c)), data provided by Google shows that in every month since April 2020, a large majority ([redacted]%) of UK users have selected Google Search as their default when presented with the Android choice screen.[38]

These results are consistent with users selecting a preferred service when presented with alternatives. A choice screen that does not materially affect user preferences may therefore add friction without improving contestability.

Evidence from earlier remedies is similar. Retrospective analysis of the 2010 Microsoft browser ballot screen finds that Internet Explorer’s share declined during the remedy period, but comparable declines occurred globally due to the growth of Chrome and Firefox. Using non-EEA countries (the United States, Canada, and Australia) as a control group, Vásquez Duque estimates a causal effect of roughly 1.4 to 2 percentage points.[39]

Overall, the empirical literature does not show that choice screens reliably transform market outcomes. As Vásquez Duque concludes in more recent work, observed effects appear to reach a ceiling, suggesting that ‘choice screens do not meaningfully alter users’ preferences’.[40]

B. Behavioural Economics and Repeated Choice Screens

The CMA proposes requiring choice screens not only at device set-up but ‘at regular points thereafter’. The rationale draws on behavioural economics. But the proposal does not sufficiently consider related concepts, including rational apathy and status-quo efficiency.

Research on behavioural interventions (‘nudges’) indicates that they can influence behaviour in low-stakes or uncertain settings but are less effective where users hold stable preferences for an experience good, such as a search engine.[41] As Vásquez Duque explains:

A rational user would search for an option as long as the alternative’s expected benefit is higher than the search costs, including the user’s time. But for many if not most users, it may make sense to stick to an option that meets a satisfactory level of quality. If this were the case, any option that met such a satisfaction level would become the user’s preferred default. And this choice may form a habit, which is likely to persist until the user experiences negative feedback or a more attractive option is brought to her attention.[42]

For many users, the marginal benefit from switching search providers may be small, while switching costs—learning a new interface or losing personalisation history—remain non-zero. Although the CMA estimates the value of users’ time,[43] the analysis gives limited attention to established research on mandated disclosure, choice fatigue, and banner blindness.[44] The consultation addresses these concerns briefly:

Google has raised a concern that the repeated display of choice screens leads to user fatigue. However, we do not consider that a relatively short prompt to consider their search choice once a year is too onerous for users. [45]

Evidence on habituation suggests that repeated prompts may not prompt careful deliberation. Users frequently respond by selecting the most familiar option simply to remove an interruption.[46] Increasing the frequency of prompts may therefore not expand effective choice but instead increase friction for users who already have settled preferences.

Recent meta-analyses of nudge interventions also indicate that effects observed in controlled or pilot settings often diminish when implemented at scale, partly because of publication bias and contextual variation.[47] Even where behavioural policies appear effective, results require cautious interpretation.

The aggregate burden is also relevant. If the UK has roughly 50 million adult mobile users and approximately 90 per cent are satisfied with Google Search, an annual mandatory choice prompt would interrupt about 45 million users who have no intention of switching, in order to reach a smaller group who might otherwise use existing device settings. This ‘time tax’ should be weighed in the proportionality assessment.

Requiring all users to make an annual active choice would therefore impose recurring costs on many users who are already satisfied with their service. The measure may address a perceived competitive concern without clear evidence of corresponding consumer harm, and its effectiveness as an instrument for altering market outcomes remains uncertain.

IV. Assessment of the Proposed Fair Ranking Conduct Requirement

The proposed Fair Ranking Conduct Requirement (Fair Ranking CR)[48] seeks to ensure that Google’s ranking decisions are ‘objective’, ‘non-discriminatory’, and ‘transparent’, supported by a publisher complaints mechanism and reporting obligations concerning ranking policies that may affect other markets.[49] These objectives are understandable. The design of the intervention, however, raises concerns about its implications for innovation and consumer welfare.

The CMA’s investigation has not identified direct evidence that individual ranking decisions are unfair. As the consultation explains:

We have not seen direct evidence that Google’s individual ranking decisions are unfair. However, taken in the round: the role of Google’s general search as a critically important digital tool for people and businesses; the lack of trust and perception of unfairness in Google’s ranking; the lack of sufficient transparency about how Google implements and operationalises its ranking in practice; and the direct impact this lack of trust has had on publishers, including deterring investment, leads us to consider that there is merit in introducing a formal requirement.[50]

The proposal therefore responds primarily to perceptions of potential unfairness and their effects on publisher behaviour.[51] This context underscores the need for caution in imposing obligations on a product whose core function is selective ranking.

The requirement of ‘objectivity’ risks misunderstanding how search operates. As Manne and Wright explain, the value of a search engine derives from its ability to prioritise some results over others.[52] Search is an exercise in curation: from billions of web pages, the system promotes a small number of results based on predicted relevance. Because ranking is inherently relative—elevating one result necessarily lowers another—differentiation is not itself evidence of anticompetitive conduct, but a necessary feature of the product.[53]

The non-discrimination obligation also presents practical enforcement difficulties. The proposal would prohibit Google from considering, among other things, whether a publisher advertises with Google, enters commercial arrangements, or exercises statutory rights. Modern search ranking, however, depends on the interaction of thousands of signals processed through machine-learning systems. Determining whether any single prohibited factor affected a particular outcome will often be infeasible. A publisher experiencing reduced traffic after a dispute may infer retaliation, but the change may equally reflect an algorithmic update, shifting user behaviour, or changes in content quality. The result could be persistent and difficult-to-resolve disputes.

Related concerns arise from the requirement that Google apply the same criteria to its own services and those of third parties. Research frequently finds that vertical integration on digital platforms produces efficiencies.[54] Integrated features—such as map results or flight modules—can reduce search costs by allowing users to obtain information directly within the search interface. As Robert Bork and Gregory Sidak observe, the consumer-welfare benefits of integration arise precisely from the immediate availability of the integrated result.[55] A categorical restriction on self-preferencing would risk sacrificing these efficiencies in order to preserve competitor distinctiveness.[56] The CMA’s interpretative notes recognise that designing a feature containing only Google inputs is not itself a breach, but its placement in rankings may be.[57] In practice, placement and design are closely connected: the usefulness of an integrated feature depends on where it appears.

The Fair Ranking CR therefore risks constraining product design in ways that reduce functionality. By attempting to mandate neutrality in a process defined by curation, and by treating integration as presumptively problematic, the proposal may degrade the user experience without clear gains in market contestability.

V. Conclusion

The CMA’s proposed Conduct Requirements for Google’s general search services constitute a significant intervention in a rapidly evolving sector. ICLE does not question the Authority’s mandate to promote competitive and contestable digital markets, nor the relevance of Google’s position in general search within the DMCC framework. As currently drafted, however, the three CRs risk imposing durable costs on UK consumers and on innovation in digital and AI services.

Across the proposals, a common concern is calibration. The measures would address uncertain or indirect harms with obligations that carry clear and immediate effects: degraded product functionality, fragmented search results, recurring user friction, and altered competitive conditions between firms subject to the regime and those outside it. The Publisher CR risks creating asymmetric regulation in generative-AI markets and using competition law to resolve copyright disputes. The User Choice CR relies on behavioural assumptions not strongly supported by real-world evidence and may impose recurring time costs on users without materially improving contestability. The Fair Ranking CR seeks to regulate ranking neutrality in a process defined by necessary curation and integration, creating practical enforcement difficulties and potential reductions in product quality.

These concerns reflect the broader insight of the error-cost framework. In dynamic markets characterised by uncertainty, the costs of mistaken intervention may be persistent, while some competitive harms may diminish through entry, innovation, and technological change. Proportionality therefore requires particular caution where conduct may benefit consumers, even if it disadvantages rivals.

ICLE respectfully encourages the CMA to anchor its final measures in demonstrable consumer harm, to rely primarily on real-world evidence, and to prefer less distortive tools—particularly transparency and attribution—over restrictions on product design. A framework that protects competition rather than competitors, and that recognises the limits of regulatory knowledge in complex markets, will better advance the DMCC Act’s objective of improving outcomes for UK consumers.

[1] Frank H. Easterbrook, The Limits of Antitrust, 63 Tex. L. Rev. 1, 14–15 (1984). Easterbrook offers the clearest exposition of the error-cost approach to antitrust, arguing that ‘[t]he economic system corrects monopoly more readily than it corrects judicial errors’.

[2] Id. at 15.

[3] Geoffrey A. Manne & Joshua D. Wright, Innovation and the Limits of Antitrust, 6 J. Competition L. & Econ. 153, 165 (2010) (extending Easterbrook’s error-cost framework to innovative markets and arguing that ‘antitrust scrutiny of innovation and innovative business practices is likely to be biased toward assigning a higher likelihood that a given practice is anticompetitive than later literature and evidence ultimately justify’).

[4] Geoffrey A. Manne, Error Costs in Digital Markets, in The Global Antitrust Institute Report on the Digital Economy 3 (Joshua D. Wright & Douglas H. Ginsburg eds., 2020) (observing that ‘[t]he risk of error is always present given the limits of knowledge, but it is magnified by the precedential nature of judicial decisions’, and that this problem is ‘further magnified when antitrust decisions address innovative, fast-moving, poorly understood, or novel market settings’).

[5] See John M. Yun, The Folly of AI Regulation, in Artificial Intelligence and Competition Policy 247, 252 (Alden Abbott & Thibault Schrepel eds., 2024) (‘Let us start, in period 0, with T = 100 and an annual growth of 30%. Due to compounding, after 10 years T grows to nearly 14× its original size. If the growth rate instead falls marginally to 25%, T grows to slightly over 9× its original size over the same period — still substantial, but about 5× lower than under 30% growth. Even a “modest” reduction in the growth rate of an emerging technology — e.g. 5% in absolute terms — can therefore produce significant long-run social-welfare losses, magnified over longer horizons’).

[6] See Geoffrey A. Manne & Dirk Auer, Antitrust Dystopia and Antitrust Nostalgia: Alarmist Theories of Harm in Digital Markets and Their Origins, 28 Geo. Mason L. Rev. 1281, 1343–45 (2021) (reviewing economic evidence on data-related theories of harm and finding incumbent data advantages far less pronounced than commonly assumed); Geoffrey A. Manne & Dirk Auer, From Data Myths to Data Reality: What Generative AI Can Tell Us About Competition Policy (and Vice Versa), CPI Antitrust Chron. (February 2024), at 12 (arguing that ‘competition or regulatory intervention to “correct” data barriers and data network and scale effects is liable to do more harm than good’).

[7] See F.A. Hayek, The Use of Knowledge in Society, 35 Am. Econ. Rev. 519, 526–27 (1945) (explaining that the price system communicates dispersed information unavailable to any single agent or planner). For an application in the competition-policy context, see Cento Veljanovski, Hayekian Competition Policy: A Historical Perspective, GW Competition & Innovation Lab Working Paper (2024).

[8] See Lazar Radic, Geoffrey A. Manne & Dirk Auer, Digital Competition Regulation: Costs, Tradeoffs, and Consequences, Int’l Ctr. for L. & Econ. (2025) (arguing that digital-competition regulation’s ‘true objectives align more with redistributing economic power, protecting less efficient competitors, and diminishing the competitive advantages of dominant digital platforms’ than with protecting consumer welfare).

[9] Brown Shoe Co. v. United States, 370 U.S. 294, 344 (1962) (‘It is competition, not competitors, which the [Sherman] Act protects’). See also Brunswick Corp. v. Pueblo Bowl-O-Mat, Inc., 429 U.S. 477, 488 (1977) (holding that plaintiffs must prove ‘antitrust injury — injury of the type the antitrust laws were intended to prevent’).

[10] See Geoffrey A. Manne & Joshua D. Wright, Google and the Limits of Antitrust: The Case Against the Case Against Google, 34 Harv. J.L. & Pub. Pol’y 171 (2011) (arguing antitrust should not infer consumer harm from conduct that disadvantages rivals, stressing the risk of false positives and the need for concrete evidence of consumer harm; identifying procompetitive rationales and welfare-enhancing design choices in search, and cautioning against condemning conduct merely because it reallocates traffic among competitors). See also Geoffrey A. Manne & Joshua D. Wright, If Search Neutrality Is the Answer, What’s the Question?, 2012 Colum. Bus. L. Rev. 151, 155 (2012) (arguing that ‘search bias’ or editorial discretion often benefits consumers by reducing users’ search costs and improving the search experience).

[11] See Robert H. Bork & J. Gregory Sidak, What Does the Chicago School Teach About Internet Search and the Antitrust Treatment of Google?, 9 J. Competition L. & Econ. 663, 680 (2013) (discussing consumer-welfare benefits of product integration in search).

[12] See, e.g., Brian Albrecht & Geoffrey A. Manne, Self-Preferencing Isn’t a Sin. It’s Often the Way Competition Works., Truth on the Mkt. (20 August 2025), https://truthonthemarket.com/2025/08/20/self-preferencing-isnt-a-sin-its-often-the-way-competition-works (explaining that self-preferencing often reflects technical or efficiency considerations rather than exclusionary intent); see also Juliette Caminade, Juan Carvajal & Christopher R. Knittel, An Economic Analysis of the Self-Preferencing Debate, 32 Competition 1 (2022) (reviewing theoretical and empirical literature on self-preferencing and dual-mode platforms and concluding the evidence is ‘mixed’, warranting a careful, case-by-case approach); see also Austl. Competition & Consumer Comm’n, Digital Platform Services Inquiry—Interim Report No. 5: Regulatory Reform 94 (September 2022), https://www.accc.gov.au/system/files/Digital%20platform%20services%20inquiry.pdf (‘Although self-preferencing conduct is often benign, conduct that leverages market power over a key online service into a related service, without a procompetitive rationale, can distort competition and decrease consumer welfare’).

[13] Competition & Mkts. Auth. (CMA), Consultation: Publisher Conduct Requirement—Google’s General Search Services (28 January 2026) (UK) [hereinafter Publisher CR Consultation].

 

[14] Id., ¶¶ 1.5, 1.10–1.11.

[15] Id. ¶ 1.5(a) (‘Given Google’s SMS in general search services, publishers have no realistic option but to allow their content to be crawled’).

[16] Id., ¶¶ 4.6–4.12 (describing the covered use cases and distinguishing grounding of search generative-AI features from training or grounding of broader generative-AI services).

[17] See Lazar Radic, Geoffrey A. Manne & Dirk Auer, Digital Competition Regulation: Costs, Tradeoffs, and Consequences, Int’l Ctr. for L. & Econ. (2025) (documenting how asymmetric digital-competition regulation can produce perverse competitive outcomes that harm consumers it aims to protect).

[18] See Manne & Auer, Antitrust Dystopia and Antitrust Nostalgia, supra note 6.

[19] See Geoffrey A. Manne, Dirk Auer, Kristian Stout, Lazar Radic & Mario A. Zúñiga, ICLE Comments to DOJ on Promoting Competition in Artificial Intelligence, Int’l Ctr. for L. & Econ. (15 July 2024), at 5–12. See also Geoffrey A. Manne, Dirk Auer & Mario A. Zúñiga, ICLE Comments to UK Competition and Markets Authority on AI Partnerships, Int’l Ctr. for L. & Econ. (9 May 2024), https://laweconcenter.org/resources/icle-comments-to-uk-competition-and-markets-authority-on-ai-partnerships.

[20] Dirk Auer & Mario A. Zúñiga, AI Partnerships and Competition: Damned if You Buy, Damned if You Don’t, ICLE White Paper 2025-08-19, at 4–5 (2025) (finding that AI partnerships are ‘largely benign from a competition-law perspective’ and that ‘no enforcement body has found concrete evidence of anticompetitive harm’ arising from them).

[21] Written Testimony of Dirk Auer, Director of Competition Policy, Int’l Ctr. for L. & Econ., Before the Subcomm. on Antitrust, Commercial & Admin. Law of the H. Comm. on the Judiciary, U.S. House of Representatives (16 December 2025) (documenting that DMA compliance forced Google to remove integrated features from European search results, producing ‘a slower, more fragmented experience’ without measurable competitive benefits).

[22] Giuseppe Colangelo, A Competition Policy Analysis of Copyright Protection in Generative AI, Sing. J.L. Stud. 271 (2025). See ICLE Comments on Artificial Intelligence and Copyright, Int’l Ctr. for L. & Econ. (30 October 2023) (analysing fair-use implications of AI training and concluding the issue is better addressed through uniform intellectual-property adjudication than firm-specific competition regulation).

[23] Publisher CR Consultation, supra note 13, ¶¶ 4.10(c)–(d) (summarising Google’s submissions that fine-tuning ‘helps the model learn how to process information rather than what current information to display’ and that allowing publishers to opt out of fine-tuning would ‘raise the risk of downranking or mis-ranking publisher content in organic search results’).

[24] See Manne & Wright, If Search Neutrality Is the Answer, What’s the Question?, supra note 10.

[25] See Manne, supra note 4.

[26] Competition & Mkts. Auth. (CMA), Consultation: User Choice Conduct Requirement—Google’s General Search Services (28 January 2026) (UK) [hereinafter User Choice CR Consultation].

[27] Id. at 52.

[28] Id. ¶ 4.4.

[29] Mozilla, Can Browser Choice Screens Be Effective? Experimental Analysis of the Impact of Their Design, Content and Placement (September 2023), https://research.mozilla.org/files/2023/09/Can-browser-choice-screens-be-effective_-Mozilla-experiment-report.pdf.

[30] Bureau Européen des Unions de Consommateurs (BEUC), An Effective Choice Screen Under the Digital Markets Act: BEUC Recommendations, BEUC-X-2023-131, at 10 (19 October 2023).

[31] Omar Vásquez Duque, The Magical Number 2 (Minus Two): An Empirical Analysis on the Efficacy of Choice Screens to Increase Competition in Digital Markets, at 17 (15 January 2026) (unpublished manuscript), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5264993.

[32] User Choice CR Consultation, supra note 26, ¶ 4.5 (‘[F]ocusing solely on levels of switching to alternative providers may be a misleading measure of a choice screen’s effectiveness, given that it should allow consumers to find a provider in line with their preferences, which may result in them staying with their existing (or incumbent) provider’).

[33] See Vásquez Duque, The Magical Number 2, supra note 31, at 3 (‘When choice screens have affected a dominant actor’s market share at all, the effect size has been, at most, 2%’).

[34] Comm’n Decision 2019/C 402/08, Case AT.40099—Google Android, 2019 O.J. (C 402) 19.

[35] Francesco Decarolis, Muxin Li & Filippo Paternollo, Competition and Defaults in Online Search, 17 Am. Econ. J.: Microeconomics 369 (2025).

[36] The study found larger market-share shifts in Russia and Turkey (>10%), driven primarily by the presence of a strong local incumbent (Yandex) able to compete on quality, rather than by the choice screen itself. See id. at 389–93.

[37] Statista, Market Share of Leading Search Engines in Europe from January 2015 to May 2025, https://www.statista.com/statistics/1386805/search-engines-market-share-all-devices-europe (last visited 16 February 2026).

[38] Competition & Mkts. Auth., Strategic Market Status Investigation into Google’s General Search Services: Final Decision ¶ 5.173(c) (10 October 2025) (UK).

[39] Omar Vásquez Duque, Active Choice vs. Inertia? An Exploratory Assessment of the European Microsoft Case’s Choice Screen, 19 J. Competition L. & Econ. 60 (2023).

[40] See Vásquez Duque, supra note 31, at 17.

[41] See Vásquez Duque, supra note 39, at 77–78.

[42] Id. at 78.

[43] User Choice CR Consultation, supra note 26, ¶ 5.14 (‘When combined with an estimate of the average value of a UK consumer’s time, this gives an average time cost per showing of the choice screen of a little under 14p’).

[44] Omri Ben-Shahar & Carl E. Schneider, The Failure of Mandated Disclosure, 159 U. Pa. L. Rev. 647, 651 (2011) (arguing that ‘mandated disclosure not only fails to achieve its stated goal but also leads to unintended consequences that often harm the very people it intends to serve’).

[45] User Choice CR Consultation, supra note 26, ¶ 4.34.

[46] Rainer Böhme & Stefan Köpsell, Trained to Accept? A Field Experiment on Consent Dialogs, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems 2403 (2010) (finding that interruption dialogs foster habituation and heuristic responses, as users are ‘trained’ to dismiss them and consent becomes increasingly ‘blind’).

[47] Bo Hu et al., Assessing Nudge Impact: A Comprehensive Second-Order Meta-Analysis, 38 J. Behav. Decision Making, no. 5, art. e70053 (2025).

[48] Competition & Mkts. Auth. (CMA), Consultation: Fair Ranking Conduct Requirement—Google’s General Search Services (28 January 2026) (UK) [hereinafter Fair Ranking CR Consultation].

[49] Id. ¶ 2.1.

[50] Id. ¶ 1.11.

[51] Id. ¶¶ 1.9–1.10.

[52] See Manne & Wright, If Search Neutrality Is the Answer, What’s the Question, supra note 10.

[53] Id. at 158.

[54] See Manne & Wright, Google and the Limits of Antitrust, supra note 10.

[55] See Bork & Sidak, supra note 11.

[56] Manne & Wright, If Search Neutrality Is the Answer, What’s the Question, supra note 10, at 189.

[57] Fair Ranking CR Consultation, supra note 48, interpretative note 5 (‘[T]he fact that a search feature (e.g. the Flights Module) might be designed and presented to include only Google inputs would not be relevant to paragraph 4.b. of the conduct requirement, but Google’s decision about where that search feature is ranked on the page would be’).

Michael Calabrese, representing New America’s Open Technology Institute (OTI), and Kristian Stout, representing the International Center for Law & Economics (ICLE), spoke via Zoom on February 19 with Jay Schwarz, Chief of the Space Bureau, with respect to the above-captioned proceeding.

We discussed the report of the LEO Satellite Policy Working Group that we jointly convened throughout 2025,[1] noting its ongoing relevance to a wide variety of work at the Space Bureau and encouraging staff to rely on it as a resource in the above-captioned and related proceedings. We expressed our view that the Working Group’s recommendations are well aligned with the goals of the above-captioned proceeding and encouraged the Commission to continue its efforts to establish a regulatory framework that will allow the nascent satellite-broadband industry to thrive.

We discussed the relevance of the Working Group’s recommendations to the Commission’s pending EPFD proceeding, particularly in relation to WRC-27. We also discussed WRC-27 Agenda Item 1.5, which will consider regulatory measures to limit unauthorized operations of non-geostationary-satellite-orbit earth stations in the fixed-satellite and mobile-satellite services, in accordance with Resolution 14 (WRC-23), and Agenda Item 1.7, which will consider studies on sharing and compatibility and develop technical conditions for the use of International Mobile Telecommunications in the frequency bands 4,400–4,800 MHz and 7,125–8,400 MHz (or parts thereof) and 14.8–15.35 GHz, taking into account existing primary services operating in these and adjacent frequency bands, in accordance with Resolution 256 (WRC-23).

Finally, we noted that spectrum policy will continue to be a significant area of focus and expressed our general support for the Commission’s spectrum-abundance agenda.

Pursuant to 47 C.F.R. § 1.1206(b), this letter is being filed with your office.

[1] See Low Earth Orbit Satellites: Policies to Promote Spectrum Sharing, Foster Competition, and Close Digital Divides, LEO Policy Working Group, Int’l Ctr. for Law & Econ. & New America (Oct. 30, 2025) [hereinafter LEO Working Group Report], https://www.newamerica.org/oti/wireless-future-project/reports/leo-satellites.

Michael Calabrese and Jessica Dine, representing New America’s Open Technology Institute (OTI), and Kristian Stout, representing the International Center for Law & Economics (ICLE), met on February 11, 2026, with Arpan Sura, senior counsel and wireless legal advisor to Chairman Brendan Carr; Will Holloway, wireless legal advisor to Commissioner Olivia Trusty; and Edyael Casaperalta, wireless advisor to Commissioner Anna Gomez, with respect to the above-captioned proceedings.

In each meeting, we summarized at a high level the recommendations in the report of the LEO Satellite Policy Working Group that we jointly convened throughout 2025.[1] The Working Group’s report highlights three salient issues for policymakers: spectrum sharing and coexistence, competition, and advancing the goals of universal and ubiquitous connectivity. The Working Group emphasized that more effective spectrum-sharing and coexistence reforms can greatly expand LEO capacity, performance, and innovation.

With respect to modernizing the rules governing GSO/NGSO spectrum sharing in the Ku and Ka bands, the Working Group strongly supported modernizing EPFD limits and adopting a framework similar to that governing NGSO/NGSO sharing.[2] This framework should require all operators to coordinate in good faith and adopt default interference thresholds that are proxies for actual harmful interference. The Working Group suggested that it would be effective to combine a short-term protection criterion based on an absolute increase in unavailability, such as 0.1% unavailability, with a long-term protection criterion based on 3% degraded throughput for GSO systems using ACM. Finally, the report suggested that the Commission should at least consider sunsets for priority protection, as it did in the NGSO/NGSO sharing framework, at least for ground stations, because these can be upgraded to minimize interference risk more easily and cheaply than GSO satellites in orbit.

Studies have shown that this framework could increase LEO providers’ overall capacity by up to eight times, providing a quality and capacity increase needed as increasing numbers of consumers rely on LEO service to get online. We urged the staff to adopt an order as soon as possible, in part to demonstrate the benefits of a more robust sharing framework well in advance of WRC-27, where this issue is on the tentative agenda. Ahead of WRC-23, OTI and ICLE joined 13 other consumer, school, library, and rural advocacy organizations in a joint letter to the leaders of the U.S. delegation strongly supporting this sharing framework.[3]

With respect to the Spectrum Abundance proceeding, we summarized the Working Group’s support for a robust satellite spectrum pipeline above 12 GHz, including most of the specific bands identified in the Commission’s pending NPRM. In particular, we expressed strong support for opening the 12.7-13.25 GHz band as an extension of the adjacent Ku downlink, as well as the 51.4-52.4 GHz band, which is also adjacent to prime FSS spectrum and unused, for NGSO uplink operations (earth to space). We also described the Working Group’s support for authorizing coordination of FSS earth stations in the 42-42.5 GHz band as part of a light-licensing and automated database coordination process. This automated coordination system could be similar to the proven 70/80/90 GHz framework and could be used to coordinate the siting of earth-station gateways in many other millimeter-wave bands, including the Lower 37 GHz band and the UMFUS bands being considered in a separate proceeding.[4]

Finally, with respect to the Space Modernization proceeding, we provided a brief overview of the comments we filed summarizing the Working Group’s recommendations.[5] The Working Group emphasized the need to replace the current process—characterized by case-by-case, bespoke reviews and tailored conditions—with standardized operational rules (including rules addressing space sustainability) capable of streamlining the application process by serving as presumed acceptable criteria. The Working Group also supported greater latitude for operators to design and modify their systems as technology continues to evolve rapidly. The Working Group recommended a “shot clock” (e.g., one year), but with flexibility for staff to pause the clock as the need for additional information or other special circumstances requires. Finally, while the Working Group did not make a recommendation on the overall timeline for deployment milestones—or whether to level the playing field between U.S. and non-U.S. licensees by mirroring the ITU timeline—it did recommend that deployment milestones be restructured into more graduated, measurable steps (e.g., every two years after an initial period), making forfeiture of larger performance bonds a stronger incentive to deter speculative applications and warehousing.

Pursuant to Section 1.1206(b) of the Commission’s rules, this letter is being filed in the above-captioned proceedings.

[1] See Low Earth Orbit Satellites: Policies to Promote Spectrum Sharing, Foster Competition, and Close Digital Divides, LEO Policy Working Group, Int’l Ctr. for L. & Econ. & New Am. (Oct. 30, 2025), https://www.newamerica.org/oti/wireless-future-project/reports/leo-satellites.

[2] See also Comments of Public Knowledge & Open Tech. Inst., Modernizing Spectrum Sharing for Satellite Broadband, SB Docket No. 25-157 (July 28, 2025).

[3] Letter from 15 Pub. Int. Orgs. to Jessica Rosenworcel, Chairwoman, FCC, Alan Davidson, Assistant Sec’y of Commc’ns & Info., Nat’l Telecomms. & Info. Admin., & Nathan C. Fick, Ambassador at Large, U.S. Dep’t of State (Aug. 28, 2023) (urging the United States to prioritize more efficient and equitable access to shared spectrum resources for both LEO and GSO networks at WRC-23).

[4] We also noted that, when considering proposals such as light licensing in the UMFUS bands, it is important to consider the reasonable investment-backed expectations of existing license holders.

[5] See Comments of Int’l Ctr. for L. & Econ. & Open Tech. Inst. at New Am., Space Modernization for the 21st Century, SB Docket No. 25-306 (Jan. 20, 2026).

 

I. Introduction and Overview

The International Center for Law & Economics (ICLE) submits these comments in response to the Federal Communications Commission’s (FCC) notice of proposed rulemaking (NPRM) on the Upper C-Band (3.98–4.2 GHz).[1] ICLE is a nonprofit, nonpartisan research center that promotes law & economics approaches to public policy. Its work seeks to ensure that regulation rests on sound economic analysis and promotes consumer welfare, particularly in dynamic communications markets.

As the FCC evaluates the record, ICLE offers an economic framework for decision-making. Broadband deployment involves substantial sunk costs and operational risks that can deter investment and leave consumers worse off. Spectrum policy should therefore reduce transaction costs, clarify operating rights, and rely on market mechanisms to allocate spectrum efficiently.

Consistent rules across the C-band are central to that objective. Aligning upper and lower C-band service rules will create scale economies in equipment; support wide, contiguous channels; and allow providers to leverage existing infrastructure investments. The Commission should also apply its emerging-technologies framework to reimburse incumbents and incentivize timely relocation, so that new licensees can deploy service quickly.

The FCC must also address harmful-interference concerns. A risk-based interference assessment—evaluating both the probability and severity of interference—will better support coexistence than worst-case assumptions that unnecessarily foreclose productive uses. That framework can accommodate flexible-use licensing while allowing compatible satellite, aviation, and shared-use operations where risk remains acceptable.

Auction design should reinforce, not undermine, these goals. The Commission should emphasize rapid clearing and post-auction flexibility in secondary markets rather than regulatory preferences embedded in expanded bidding credits, allowing price discovery to identify the highest-valued uses.

Finally, close coordination with the Federal Aviation Administration (FAA) and other federal agencies is essential to avoid the delays experienced in the lower C-band transition. Transparent, predictable interagency processes will reduce regulatory uncertainty, support investment, and enable timely deployment of advanced wireless services.

II. Auction Design and Transition Policy for the Upper C-Band

The record supports the FCC’s proposal to adopt the same service rules for the upper C-band as for the lower C-band. The lower C-band framework generated substantial value for licensees and sufficient auction revenues to fund incumbent relocation, while a unified C-band approach increases technical efficiency, lowers deployment costs, and improves investment certainty.

The Commission can also apply its existing emerging-technologies framework to the upper C-band transition. Using auction proceeds and relocation incentives aligns incumbent behavior with rapid clearing and reduces coordination delays, enabling new licensees to deploy service sooner and capture the broader social value of the spectrum.

Consistent rules and timely clearing further support wide, contiguous channels, equipment standardization, and economies of scale, all of which strengthen network performance and accelerate deployment. The Commission should therefore prioritize rapid relocation, predictable operating rules, and post-auction flexibility in secondary markets, while avoiding auction design choices—such as expanded bidding credits—that substitute regulatory preferences for market-based price discovery.

A. Harmonizing Upper and Lower C-Band Rules to Reduce Risk and Deployment Costs

The Commission’s spectrum policy should maximize social welfare by putting radio spectrum to its highest-valued use. Doing so requires more than successful auctions. Licensees must make large, forward-looking investments to deploy service, and those investments carry substantial risk.[2]

Wireless deployment is capital-intensive. Providers must obtain construction permits and zoning approvals, secure access to public rights-of-way, install backhaul, and satisfy environmental and historic-preservation review before earning any return on the spectrum asset.[3] These sunk and uncertain costs shape expected returns. Because the upper C-band will largely support 5G networks, providers also face intense competition from rival networks, which constrains margins and limits upside.

As deployment costs and risks rise, the expected value of licenses falls. That reduces both bidders’ willingness to pay at auction and their incentives to deploy quickly and broadly after the auction. One way to mitigate these effects is to lower transaction costs. The Commission has pursued multiple proceedings aimed at doing so, but it cannot eliminate them entirely.

At its core, spectrum management minimizes the transaction costs of radio coordination. Harmful interference threatens service quality, customer relationships, and brand reputation. As interference risk rises, so does the risk to network investment. Identifying and resolving interference, though, also imposes transaction costs.

The FCC reduces these costs by defining license rights clearly and by providing information about who operates where. Still, it cannot eliminate interference risk. Residual risk increases uncertainty about reliability, service quality, and returns on invested capital. Spectrum policy therefore should provide clear, predictable operating rules and enable scale efficiencies that improve risk-adjusted returns for large sunk investments.

For this reason, the Commission should adopt the same service rules and band plan in the upper C-band as in the lower C-band. Power levels, guard-band alignment, emission masks, and antenna rules should match so that 3.7–4.2 GHz functions as a single, continuous operating environment using the same network equipment.[4]

First, harmonized rules allow wide, contiguous 100 MHz channels. One 100 MHz channel is more spectrally efficient than five 20 MHz channels, because each channel requires guard bands and control signaling that does not carry user traffic.[5] Combining fragmented spectrum eliminates these duplicative inefficiencies.

Carrier aggregation can combine nonadjacent bands, but it does not deliver the same benefits as contiguous spectrum.[6] Aggregation requires smartphones to monitor multiple frequencies and process simultaneous data streams, increasing power consumption and battery drain.[7] Not all devices support complex aggregation combinations, so lower-cost devices cannot realize wider bandwidth benefits without contiguous spectrum. Multiple signals also introduce small scheduling delays that increase latency and can limit ultra-low-latency applications.[8]

Second, harmonization lets providers reuse lower C-band equipment and reduce deployment costs. Carriers have already upgraded thousands of towers with lower C-band radios, which required tower reinforcement, new fiber and power runs, and installation of massive-MIMO equipment. Modern radios often support wide frequency ranges, typically 3.7–4.2 GHz and sometimes 3.3–4.2 GHz.[9] If the rules match, providers can activate upper C-band service through software updates at near-zero incremental installation cost.

Even where new radios are required, carriers can deploy a single wideband unit covering both bands, instead of leasing tower space for separate equipment. Divergent rules—such as different power limits—would force installation of additional radios and create direct and indirect costs.

Harmonization reduces duplicative expenses and deployment risk. That produces direct effects (greater deployment), indirect effects (earlier and more efficient use of construction, engineering, and integration supply chains), and induced effects (productivity gains from faster, lower-latency connectivity). One study estimates each additional 100 MHz of spectrum contributes $264 billion in GDP, 1.55 million jobs, and $388 billion in consumer surplus.[10]

Some commenters urge the Commission to auction smaller channels or narrower geographic licenses to give smaller bidders more opportunity to compete with nationwide providers.[11] Larger licenses can leave some areas initially underutilized, but that critique overlooks their transaction-cost advantages at initial assignment. Efficient allocation often fails when parties must assemble rights across many licenses, markets, and channels.

By assigning broader, clearly defined rights at the outset, the Commission reduces assembly costs and increases auction value. After assignment, licensees can partition, disaggregate, lease, or sell spectrum access where efficient. Secondary markets then allocate spectrum to higher-valued uses over time.

Accordingly, the Commission should not attempt to optimize spectrum use by fragmenting licenses ex ante. Instead, it should maximize post-auction flexibility.[12] The FCC can do so by revisiting partitioning and disaggregation rules and lowering transaction costs for licensees that transfer rights across markets or channels. Secondary markets, rather than regulatory fragmentation, provide the most efficient mechanism to allocate operating rights.

B. Incentivizing Rapid Incumbent Relocation Through the Emerging-Technologies Framework

Wireless providers have strong incentives to deploy service in the upper C-band. Still, reallocation and deployment risks can materially reduce license values relative to the broader economic gains deployment creates. FCC policy therefore should prioritize clearing existing operations quickly, rather than maximizing net auction revenue. The social value of a rapid transition likely exceeds the Treasury’s share of auction proceeds.

The Commission’s emerging-technologies framework, applied in the lower C-band, sought to maximize the value of reallocation by using auction proceeds to incentivize incumbents to relocate. The framework reimbursed necessary relocation costs and offered accelerated-relocation payments before new licensees could begin operations, while also providing time for incumbents to move their services.[13]

Relocation carries operational risk. Incumbents will not accelerate when the risk of service disruption exceeds the compensation offered. Mandatory deadlines alone invite delay, extension requests, and resistance when incentives misalign with rapid clearing. Incentive payments instead align private incentives with the Commission’s objective of timely relocation, reducing delay and coordination costs.

The framework also supports full clearing of the band. Complete clearing both increases auction value and provides technical flexibility for new licensees, creating a greenfield environment for new services. It also aligns the U.S. band plan with the International Mobile Telecommunications (IMT) standards issued globally by the International Telecommunication Union’s (ITU) ITU Radiocommunication Sector (ITU-R), enabling equipment scale economies and lowering deployment costs.

Applying the framework to the upper C-band may require adjustments. Some commenters note that relocation may require alternative transmission approaches. For example, NCTA argues reimbursement should cover not only capital expenditures, but also new operating expenses,[14] and that incumbents need adequate time to maintain service quality and reliability.[15] If incumbents face higher operating costs after transition, relocation will slow unless those costs are addressed. To promote speed, the FCC should allow compensation beyond capital investment where necessary.

The Commission’s objective should be rapid relocation, not Treasury revenue. When applying the emerging-technologies framework to the upper C-band, the FCC should provide sufficient transition funding to reduce relocation risk and enable efficient migration to alternative technologies or frequency bands. Although this approach may reduce net auction proceeds, it will shorten deployment timelines, reduce coordination costs, and allow consumers and communities to realize the benefits of high-speed connectivity sooner.

C. Limiting Bidding Credits to Preserve Market-Based Price Discovery

Some commenters urge the Commission to expand bidding credits beyond those in 47 C.F.R. § 1.2110(f)(2)(i)(B)–(C). Most notably, WISPA seeks a 35% credit for businesses with average gross revenues of $4 million or less over the preceding five years.[16] The argument is that smaller bidders lack the financial resources to compete with nationwide 5G providers but could deliver greater consumer value.

As discussed above, the social value of spectrum auctions often exceeds auction proceeds. Incentivizing incumbents to relocate does not require the Commission to determine which service is more valuable. Relocation payments instead facilitate market price discovery by allowing bidders to reveal how much they value the spectrum.

Bidding credits are different. They require the Commission to favor certain bidders or business models over others. History shows the difficulty of such determinations.[17] Regulators face limited information, potential capture, and incentive misalignment. For these reasons, the FCC has increasingly relied on market mechanisms to determine which users and services value spectrum most highly.

Although the Commission proposes 15% and 25% bidding credits, it should eliminate bidding credits to the extent the Communications Act permits.[18] Removing them would avoid regulatory value judgments and allow the market to determine the highest-valued use of the licenses.

If the Commission retains credits, it should limit them to the proposed 15% and 25% levels and align the auction design with the lower C-band auction.[19]

III. Risk-Based Interference Management and Spectrum Sharing

As a threshold matter, the Commission should maximize concurrent operations in the upper C-band so long as those operations do not cause harmful interference to neighboring services. Historically, interference analysis relied on worst-case assumptions.[20] Regulators evaluated whether interference was physically possible under extreme operating parameters.[21] This approach effectively prevents harmful interference, but it also imposes significant opportunity costs by excluding many otherwise compatible uses.

A risk-based interference assessment takes a more economically grounded approach. Instead of asking whether interference could ever occur, the regulator evaluates how likely it is and how severe the consequences would be.[22] Using probabilistic methods, including Monte Carlo simulations, the analysis categorizes potential harms along a likelihood distribution. The regulator then determines what level of risk is acceptable, given the severity of the harm. A rare but catastrophic outcome would be unacceptable, while a more common event causing only minor service degradation may be tolerable. The key is balancing probability against consequence.

Introducing flexible-use licenses in the upper C-band will increase interference concerns, particularly regarding radio altimeters operating in the 4.2–4.4 GHz band. As the Commission evaluates the technical record and coordinates with other federal agencies, it should assess these concerns through a risk-based framework.

Several commenters also propose productive uses for portions of the band not auctioned for exclusive licenses. SpaceX asks the Commission to authorize supplemental coverage from space in the remaining spectrum.[23] It contends next-generation satellite systems could use the band for downlink and provide direct-to-device service without interfering with radio altimeters.

Similarly, Monisha Ghosh and co-authors propose allocating remaining spectrum to a shared-access regime similar to the CBRS band (3.55–3.7 GHz).[24] They explain:

[T]his left-over spectrum could be utilized in terrestrial wireless deployments by leveraging the successful sharing mechanism used in the CBRS band (3.55 – 3.7 GHz). The Spectrum Access System (SAS) developed for CBRS can be easily extended to determine appropriate power levels for new devices in this band such that no harmful interference occurs to both in-band (FSS) and adjacent band (altimeters) incumbents.[25]

Such an approach could expand use of the band and enable participation by smaller providers unable to obtain nationwide exclusive licenses.

As with terrestrial flexible-use operations, the Commission should evaluate both proposals using risk-based interference analysis. Each could be viable if operations can coexist with incumbents and new terrestrial services at acceptable risk levels. The same analytical framework should govern full-power licenses and shared or satellite operations alike.

IV. Interagency Coordination and Aviation-Sector Certainty

The record supports continued FCC coordination with the Federal Aviation Administration (FAA) and other federal agencies as the Commission evaluates terrestrial use of the upper C-band. AT&T notes that “[m]aking this spectrum available within the congressionally mandated timeframe will be challenging, and as Chairman Carr notes, will require ‘extensive cooperation and information sharing between the wireless and aviation sectors.’”[26]

If the Commission seeks to maximize the value of upper C-band reallocation, additional deployment delays will undermine that objective. Delays create uncertainty about the scope of license rights. Without clear and reliable rights, bidders cannot value spectrum accurately, resources are misallocated, and licenses may not go to their highest-valued uses.

Delays also impose indirect costs. The lower C-band coordination process has already introduced uncertainty into spectrum planning. Repeating that experience in this proceeding would discourage investment and slow deployment even before operations begin.

Effective coordination with the FAA and other federal agencies is therefore essential to mitigate risk and establish predictable operating conditions. Successful coordination would also signal that prior interagency failures have been resolved and reduce future concerns about regulatory disruption of licensed spectrum use.

The process will also require substantial aviation-industry participation, which has already begun in this proceeding. Sustained collaboration will help establish clear operating expectations and provide the certainty necessary for both wireless and aviation stakeholders.

V. Conclusion

Making the upper C-band available for flexible use presents a significant opportunity to expand U.S. leadership in 5G and next-generation wireless services. To capture the full social and economic value of this spectrum, the Commission should prioritize efficiency, certainty, and rapid deployment over short-term revenue maximization or regulatory micromanagement.

The record points to a clear path. Harmonizing the upper and lower C-band service rules will reduce transaction costs, support wide, contiguous channels, and allow providers to leverage existing infrastructure investments and standardized equipment. Applying the emerging-technologies framework to reimburse incumbents and incentivize timely relocation will accelerate clearing and enable licensees to deploy service sooner. The Commission should likewise rely on market-based auction design and post-auction flexibility, rather than expanded bidding credits, to ensure spectrum flows to its highest-valued use.

Risk-based interference assessments and sustained coordination with the FAA and other stakeholders will further provide the regulatory certainty needed for investment, while allowing compatible uses to coexist. Taken together, these policies will create a predictable, high-capacity mid-band environment that speeds deployment, lowers costs, and enables the productivity gains and innovation that widespread advanced wireless connectivity makes possible.

[1] In the Matter of Upper C-band (3.98–4.2 GHz), Notice of Proposed Rulemaking, 47 C.F.R. pts. 1, 2, 25 & 27, GN Docket No. 25-59, FCC 25-78, 90 Fed. Reg. ___ (Dec. 5, 2025), https://www.federalregister.gov/documents/2025/12/05/2025-22020/in-the-matter-of-upper-c-band-398-42-ghz [hereinafter NPRM].

[2] See Reply Comments of the International Center for Law & Economics, Build America: Eliminating Barriers to Wireless Deployment, WT Docket No. 25-276 (Jan. 12, 2026), https://laweconcenter.org/wp-content/uploads/2026/01/Build-America-Wireless-infrastructure-Comments.pdf.

[3] See Comments of the International Center for Law & Economics, CTIA Petition for Rulemaking, RM-12003 (Apr. 30, 2025).

[4] Comments of Ericsson, Upper C-band (3.98 to 4.2 GHz), GN Docket No. 25-59, at 13 (Jan. 20, 2026), https://www.fcc.gov/ecfs/document/1012043306275/1; Comments of AT&T Services, Inc., Upper C-band (3.98 to 4.2 GHz), GN Docket No. 25-59 (Jan. 20, 2026), https://www.fcc.gov/ecfs/document/101201785127405/1.

[5] CTIA, Smarter and More Efficient: How America’s Wireless Industry Maximizes Its Spectrum 7 (2019), https://api.ctia.org/wp-content/uploads/2019/07/Spectrum_Efficiency.pdf.

[6] Haythem Banu Salameh, Marwan Krunz, & David Manzi, Spectrum Bonding and Aggregation with Guard-Band Awareness in Cognitive Radio Networks, 13 IEEE Trans. Mobile Comput. 569 (2014), https://doi.org/10.1109/tmc.2013.11.

[7] Signals Research Group, Paving the Way to 5G SA with Carrier Aggregation (Feb. 2023), https://www.nokia.com/asset/f/213121 (carrier aggregation can reduce device battery use in some cases by completing downloads faster despite higher current draw).

[8] Ericsson, What, Why and How: the Power of 5G Carrier Aggregation (2021), https://www.ericsson.com/en/blog/2021/6/what-why-how-5g-carrier-aggregation.

[9] Press Release, Samsung, Samsung Introduces Complete C-band Network Solutions Portfolio (Apr. 20, 2021), https://news.samsung.com/global/samsung-introduces-complete-C-band-network-solutions-portfolio.

[10] Hector Lopez & Julien Martin, The Economic Impact of Each Additional 100 MHz of Mid-band Spectrum for Mobile (prepared for CTIA, Jan. 22, 2025), https://api.ctia.org/wp-content/uploads/2025/01/The-economic-impact-of-allocating-mid-band-spectrum-to-mobile.pdf.

[11] Comments of WISPA, Upper C-band (3.98–4.2 GHz), GN Docket No. 25-59, at 5 (Jan. 20, 2026), https://www.fcc.gov/ecfs/document/101201162526102/1.

[12] See 47 C.F.R. § 101.1111.

[13] NPRM, supra note 1 at n. 2.

[14] Comments of NCTA – The Internet & Television Association, Upper C-band (3.98–4.2 GHz), GN Docket No. 25-59 (Jan. 20, 2026), https://www.fcc.gov/ecfs/document/10121058788895/1.

[15] Id. at 18–19. (transition requires planning, satellite construction and launch, technical upgrades, relocation, alternative-path deployment, testing, and decommissioning).

[16] WISPA, supra note 11, at 3.

[17] See Use of the 5.850–5.925 GHz Band, Report and Order, Further Notice of Proposed Rulemaking, and Order of Proposed Modification, 35 FCC Rcd. 13440, ET Docket No. 19-138 (Nov. 20, 2020), https://docs.fcc.gov/public/attachments/FCC-20-164A1.pdf (the Commission originally allocated the 5.9 GHz band to dedicated short-range communications (DSRC) for vehicle safety, expecting deployment, but the technology did not develop and the band was later repurposed in part for unlicensed use and in part for another standard).

[18] 47 U.S.C. § 309(j).

[19] 47 C.F.R. § 1.2110(f)(2)(i)(B)–(C).

[20] Spectrum and Receiver Performance Working Group of the Federal Communications Commission’s Technological Advisory Council, A Quick Introduction to Risk-Informed Interference Assessment (Apr. 1, 2015), https://transition.fcc.gov/bureaus/oet/tac/tacdocs/meeting4115/Intro-to-RIA-v100.pdf.

[21] Jean Pierre De Vries, Risk-Informed Interference Assessment: A Quantitative Basis for Spectrum Allocation Decisions, 41 Telecomm. Pol’y 434 (2017), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2792395.

[22] Id.

[23] Comments of Space Exploration Holdings, LLC, Upper C-band (3.98–4.2 GHz), GN Docket No. 25-59 (Jan. 20, 2026), https://www.fcc.gov/ecfs/document/10121052152521/1.

[24] Comments of Monisha Ghosh, Joshua Roy Palathinkal, Muhammad Rochman, & S. M. Haider Ali Shuvo, Upper C-band (3.98–4.2 GHz), GN Docket No. 25-59 (Jan. 21, 2026), https://www.fcc.gov/ecfs/document/10121176025934/1.

[25] Id. at 1-2.

[26] AT&T Services, Inc., supra note 4, at 9.

As former federal antitrust enforcers, we have devoted significant portions of our careers to protecting consumers and competition and we continue to support vigorous enforcement. We are writing to encourage the Department to review the pending Netflix-Warner Bros. merger based on proven standards that evaluate deals based on their impact on consumer welfare, rather than the progressive analytic framework fabricated during the prior administration.

Specifically, we recommend the following:

• First, we urge DOJ to evaluate the merger under proven criteria, rather than the 2023 Merger Guidelines. Those Guidelines, adopted on a straight partisan vote, minimize, ignore, or dismiss the benefits of merger efficiencies, rely on outdated structural presumptions, and adopt theories of harm unrelated to established indicia of competition. By relying on these Guidelines, the Department would effectively give the prior administration a say in reviewing this deal.
• Second, based on a wealth of precedent and empirical evidence, we encourage DOJ to recognize that most mergers, particularly vertical mergers, raise no competitive concerns because they are either benign or promote competition. The Netflix-Warner Bros. merger has both vertical and horizontal elements, but the crux of the deal is vertical in nature. There is a very credible argument that the deal strengthens competition by pairing world-class content creation with global distribution, allowing a newly integrated challenger to compete more effectively in a dynamic market.
• Third, to the extent that DOJ ultimately harbors some competitive concerns, we encourage it to consider negotiating remedies tailored to address those concerns, including the possibility of behavioral remedies, rather than seeking to block the merger outright. Across both the Department and FTC, the Trump Administration has consistently negotiated reasonable remedies that allow otherwise pro-competitive deals to move forward.
• Fourth, and perhaps most importantly, we encourage DOJ to avoid reliance on speculative theories, such as those based on notions of foreclosure, potential competition, or structural presumptions, all of which were embraced uncritically by the prior administration. If the evidence shows that a merger would harm consumers, it should be modified to address the consumer harm, and only if that is not possible, Merger review, however, should not be used to attack companies or engineer bureaucratic hurdles to economic freedom, progress, and growth, particularly when, as here, a merger would improve the global position of U.S. companies in critical markets.

These principles hold true for any merger that DOJ might review. We describe them more fully below.

A.     In Evaluating the Merger, DOJ Should Rely on Proven Criteria Rather than the 2023 Merger Guidelines

Adopted by a purely partisan vote during the Biden Administration, the 2023 Merger Guidelines seek to rewrite decades of antitrust policy by declaring structural presumptions against mergers that increase market concentration and by downplaying the possibility of merger efficiencies.[1] The Guidelines rely on selective and outdated cases and economic ideas while ignoring decades of economic learning and recent court decisions that reject these flawed theories.[2] The Guidelines also undermine the rule of law by affording the agencies tremendous discretion to pick winners and losers, dictate market structures, and play to favored constituencies. To date, it appears that no court has cited the Guidelines for any of its more aggressive principles.

Instead of relying on the prior administration’s framework, DOJ should evaluate the merger based on the proven bipartisan criteria of how a merger will affect consumer welfare, including prices, output, quality, variety, and innovation. In this case, credible evidence suggests that the Netflix-Warner Bros. merger may increase output by broadening the availability of Warner Bros.’ existing library and by leading to investment in new production facilities; similarly, the merger could lower prices by reducing the need for millions of consumers to subscribe to both Netflix and HBO Max. In any event, the relevant question should be whether or not the merger is likely to harm competition and consumers. DOJ should evaluate the merger based on established antitrust jurisprudence, economic learning, and the facts and circumstances presented, not “progressive” wish lists labeled as Guidelines.

B.     DOJ Should Recognize that Most Mergers, Particularly Vertical Mergers, Either Promote Competition or are Competitively Benign

In 2023, a review of the existing empirical literature found that “There is zero basis to doubt the once-settled wisdom underpinning the basic framework for merger review: mergers can and do advance procompetitive business objectives.”[3] Based on this type of evidence, former Assistant Attorney General for the Antitrust Division Christine Varney declared that “the vast majority of mergers are either procompetitive and enhance consumer welfare or are competitively benign.”[4]

Similarly, both the Department and FTC have recognized that mergers “are one means by which firms can improve their ability to compete.”[5]

In 2018, Makan Delrahim, then Assistant Attorney General for the Antitrust Division, emphasized that “most mergers are pro-competitive, or at least competitively neutral.” Quoting one of his predecessors, he stressed that “mergers are ‘an important and extremely valuable capital market phenomenon, that they are to be in general facilitated, and that it is socially desirable that uncertainty and risk be removed wherever possible to do so, subject, of course, to the very important limitation that where a merger threatens significantly to lessen competition, it should be halted.’”[6] In the same vein, the FTC’s current Chairman, Andrew Ferguson, has stated that mergers “are a critical way in which capital fuels innovation” and that an acquisition, along with new management, “can unleash new vitality, innovation, and growth.”[7]

In particular, courts have consistently recognized that vertical mergers characteristically promote competition. In FTC v. Microsoft, for example, the court stated that “many vertical mergers create vertical integration efficiencies between purchasers and sellers,” that “vertical integration creates efficiencies for consumers,” and that “Vertical integration is ubiquitous in our economy and virtually never poses a threat to competition when undertaken unilaterally and in competitive markets.”[8] This finding is consistent with the economic literature, which recognizes the efficiencies and welfare-enhancing benefits that tend to be associated with vertical acquisitions.[9]

Although this merger includes both horizontal and vertical elements, the core of the merger is vertical in nature: it will allow Netflix to distribute Warner Bros.’ existing properties to millions of new customers.

C.      If Necessary, DOJ Should Consider the Use of Tailored Remedies

To the extent that the Department ultimately harbors some competitive concerns about the merger, we encourage DOJ to negotiate tailored remedies, including the possibility of behavioral remedies, to resolve those concerns, rather than seek to block the merger outright. Across both the Department and FTC, the Trump Administration has consistently negotiated reasonable remedies that allow otherwise pro-competitive deals to move forward. For example, in Microsoft-Activision Blizzard, Omnicom-Interpublic, and Keysight Technologies-Spirent, the antitrust agencies worked closely with the merging companies to resolve their concerns, all mergers that benefited consumers and that allowed U.S. companies to improve their global competitiveness. In contrast, when the antitrust agencies fully block mergers rather than address specific concerns with tailored remedies, the agencies often reduce competition, harm consumers, and degrade the U.S. economy. For instance, the antitrust agencies’ inflexibility resulted in bankruptcies and lost jobs in both the failed Amazon-iRobot and Jet Blue-Spirit mergers. In this merger, and similar to the deals approved in Microsoft-Activision Blizzard and Amgen-Horizon Therapeutics, Netflix has already committed to licensing Warner Bros.’ properties to other platforms.

D.     DOJ Should Avoid Speculative Theories, Particularly When a Merger Would Improve the Global Position of U.S. Companies in Critical Markets

The prior administration consistently embraced speculative theories, such as those based on notions of foreclosure, potential competition, or structural presumptions, to block mergers that would have enhanced competition. The list is long: Nvidia-Arm, Lockheed-Aerojet, Meta-Within, Illumina-Grail, Amazon-iRobot, and Microsoft-Activision Blizzard. These were all mergers with vertical components that, if consummated, would have enhanced competition and improved the global position of U.S. companies in various critical markets, including chip design, rocket motors for missiles, robotics, the metaverse, cancer treatments, and the gaming industry. In other words, during the prior administration, merger review often reached a point where the antitrust agencies were pursuing speculative theories in ways that undermined our national interests in globally competitive markets.

To be clear, antitrust agencies should examine mergers. After all, antitrust law protects competition and consumers, not companies seeking to merge, nor their competitors. Our laws do not and should not give preferential treatment aimed at promoting “national champions.” If the evidence shows that a merger likely would harm consumers, it should be modified to address the consumer harm, and only if that is not possible, blocked. Merger enforcement, however, should never be used to attack companies, pick winners and losers, or engineer bureaucratic hurdles to economic freedom, progress, and growth, particularly when a merger would improve the global position of U.S. companies in critical markets.

The proposed Netflix–Warner Bros. merger holds the potential to enhance competition, as well as to improve the global competitiveness of U.S. companies in the entertainment sector.[10]

[1] U.S. Chamber, The Final Merger Guidelines: A Nightmare Before Christmas? (Dec. 19, 2023), https://www.uschamber.com/antitrust/the-final-merger-guidelines-a-nightmare-before-christmas.

[2] Jason Furman and Carl Shapiro, How Biden Can Get Antitrust Right, WSJ (July 27, 2023), https://www.wsj.com/opinion/how-biden-can-get-antitrust-right-khan-ftc-justice-department-guidelines-11364639.

[3] U.S. Chamber, Evidence of Efficiencies in Consummated Mergers (June 1, 2023), at https://www.uschamber.com/assets/documents/20230601-Merger-Efficiencies-White-Paper.pdf.

[4] Statement of Ass’t Att’y Gen. Christine Varney, Merger Guidelines Workshops, Third Annual Georgetown Law Global Antitrust Enforcement Symposium (Sept. 22, 2009).

[5] OECD, Conglomerate Effects of Mergers – Note by the United States to the Organisation for Economic Co-operation and Development (June 4, 2020) at 5, https://www.ftc.gov/system/files/attachments/us-submissions-oecd-2010-present-other-international-competition-fora/oecd-conglomerate_mergers_us_submission.pdf.

[6] Delrahim, Remarks at the 2018 Global Antitrust Enforcement Symposium (Sept. 25, 2018) (quoting William Baxter), https://www.justice.gov/archives/opa/speech/assistant-attorney-general-makan-delrahim-delivers-remarks-2018-global-antitrust.

[7] Statement of Chairman Ferguson, Joined by Commissioners Holyoak and Meador, In the Matter of Synopsys, Inc. / Ansys, Inc. (May 28, 2025), https://www.ftc.gov/system/files/ftc_gov/pdf/synopsys-ansys-ferguson-statement-joined-by-holyoak-meador.pdf.

[8] FTC. v. Microsoft Corp., No.23-cv-2880-JSC, 2023 WL 4443412 (N.D. Cal. Jul. 10, 2023) (citations omitted).

[9] See, e.g., Francine Lafontaine & Margaret Slade, Vertical Integration and Firm Boundaries: The Evidence, 45 J. ECON. LIT. 629, 677 (2007) (“In spite of the lack of unified theory, over all a fairly clear empirical picture emerges. The data appear to be telling us that efficiency considerations overwhelm anticompetitive motives in most contexts. Furthermore, even when we limit attention to natural monopolies or tight oligopolies, the evidence of anticompetitive harm is not strong.”).

[10] See Asheesh Agarwal, Netflix-Warner Bros. Merger Will Enhance America’s Global Influence, Townhall (Jan. 29, 2026), https://townhall.com/columnists/asheeshagarwal/2026/01/29/netflixwarner-bros-merger-will-enhance-americas-global-influence-n2670259.

I. Introduction

The International Center for Law & Economics (ICLE) welcomes the opportunity to comment on the public consultation initiated by the Ministry of Industry and Trade (MOIT) of the Socialist Republic of Vietnam on proposed amendments to Law No. 23/2018/QH14 of June 12, 2018, on Competition (Competition Law).[1] Vietnam has emerged as one of Southeast Asia’s most dynamic and fast-growing digital economies, supported by a regulatory approach that has balanced legal certainty with the flexibility needed for technological experimentation.[2]

The current drafting process, informed by Resolutions No. 57-NQ/TW and No. 68-NQ/TW, aims to strengthen the legal framework for the digital economy. The proposed introduction of prescriptive, ex ante prohibitions targeting digital platforms in Article 27 nonetheless risks reversing this successful trajectory. These provisions draw heavily on regulatory models from the European Union and the United Kingdom that have already produced significant shortcomings and unintended consequences.[3]

The proposed amendments focus on what MOIT characterizes as the abuse of dominant or monopolistic positions by digital platforms.[4] This approach assumes that existing competition law cannot adequately address features of digital markets, such as network effects and data advantages, or that these features inherently signal competitive harm rather than efficient competition. A law & economics perspective points in the opposite direction. Digital markets typically exhibit dynamic rivalry, in which firms compete to displace one another through innovation rather than merely to protect static market positions. As explained below, prohibitions on practices such as self-preferencing, tying, and mandated data access are likely to reduce consumer welfare, deter innovation, and weaken the security of Vietnamese users.

II. Ex Ante Platform Regulation Conflicts with Vietnam’s Dynamic Market Reality

The Party’s guidance on lawmaking, particularly Resolution No. 66-NQ/TW on reforming legislation to meet national development needs, emphasizes that legal frameworks must closely reflect real-world conditions and remain grounded in Vietnam’s specific economic and institutional context.[5] The MOIT’s proposal nonetheless closely tracks the European Union’s Digital Markets Act (DMA) and the United Kingdom’s Digital Markets, Competition and Consumers Act (DMCCA). These regimes mark a decisive departure from effects-based competition analysis focused on consumer welfare toward a more formalistic, “fairness”-oriented approach that prioritizes the protection of competitors over the competitive process itself. They also reflect political, institutional, and geopolitical priorities specific to the EU and the UK—rather than neutral economic principles—that Vietnam may not share and need not import.

Claims that digital markets require ex ante intervention rest on a static view of competition. Conventional antitrust analysis often treats stable market shares as evidence of durable market power. In digital ecosystems, however, stable shares more often reflect the temporary rewards of successful innovation than insulation from competitive pressure.[6] The constant risk of displacement by superior technologies or business models pushes even leading firms to invest continuously in research and development. Prescriptive and rigid rules in such environments risk locking business models in place and suppressing the dynamic rivalry that enables new entrants to challenge incumbents.

The MOIT’s emphasis on the “intermediary” role of digital platforms further signals concern about so-called bottleneck or gatekeeper power.[7] This framing overlooks the “Host’s Dilemma,” under which platforms must strike a careful balance between openness to third-party complementors and sufficient control to preserve security, quality, and commercial viability. When platforms succeed by offering integrated features, user demand typically reflects that success. Mandated unbundling or enforced neutrality may therefore compel firms with strategic market positions to degrade valued products and services, ultimately harming the consumers the regulation seeks to protect.

III. Self-Preferencing Is Not Presumptively Anticompetitive

Article 27(2)(a) of the MOIT’s proposal would prohibit “self-preferential treatment,” under which a platform prioritizes its own products or services through rankings, algorithms, or technical design choices.[8]  Framed this way, the proposal effectively adopts what ICLE scholars describe as a vertical-discrimination presumption—the view that vertical integration, or closely related conduct, is inherently suspect and presumptively anticompetitive absent compelling justification.[9]

This presumption conflicts with established insights from industrial organization economics. Firms often engage in vertical integration and related forms of self-preferencing to improve efficiency, reduce transaction costs, enhance product quality, enable new functionality, or support cross-subsidies that expand output. Self-preferencing can, in limited circumstances, raise competitive concerns. A categorical prohibition, by contrast, would likely condemn a broad range of conduct that benefits consumers, while doing little to address the narrower set of practices that could plausibly increase quality-adjusted prices or deter innovation.

A. The Host’s Dilemma and Dynamic Platform Design

Much of the intuition behind self-preferencing bans reflects what Jonathan Barnett describes as the “host’s dilemma.”[10] Complementors may grow dependent on a platform’s rules, distribution, and ranking systems, while the platform retains discretion to redesign its environment in ways that favor its own offerings. This dynamic is not unique to digital markets. It arises whenever firms invest under uncertainty, particularly when those investments tie closely to a specific relationship or distribution channel. Transaction-cost economics describes this condition as asset specificity: when investments carry greater value within a particular relationship than outside it, the risk of opportunism rises and governance mechanisms—such as contracts, integration, reputational constraints, or tailored rules—play a central role.[11]

Firms often manage these risks through contracts. In more arm’s-length relationships—e.g., a website that optimizes for search traffic—parties may not negotiate bespoke terms that guarantee stable rankings or interfaces. In those settings, the relevant baseline is not a right to neutrality but an expectation that platform design will evolve over time. Blanket non-discrimination mandates that freeze platform design to protect complementors can introduce their own distortions, including encouraging inefficient overinvestment in business models tailored to static platform rules, rather than to consumer value.[12]

Self-preferencing also commonly reflects standard integration and product-design choices. Coordinating complementary services within a single technical stack can reduce latency, improve reliability, and enable features that loose interoperability cannot easily deliver.[13] In cloud and data-intensive environments, performance often depends on data locality and tightly coupled scheduling. Rules that require “neutrality” by restricting integration can therefore degrade service quality and increase costs.[14] The central point is not that every integration choice is harmless, but that many reflect product improvements and cost reductions that competition policy should hesitate to prohibit categorically.

B. Self-Preferencing Frequently Benefits Consumers

Claims that self-preferencing is typically harmful find little support in the empirical literature. Across a range of platform settings, downstream entry or preferential placement of first-party offerings often coincides with market expansion, greater user awareness, and increased innovation by complementors—outcomes that conflict with a presumption of systematic foreclosure.

Empirical studies illustrate this pattern. Zhuoxin Li and Ashish Agarwal find that Facebook’s integration of Instagram increased demand not only for Instagram itself, but also for photography apps more broadly.[15] The integration raised awareness and expanded the market in ways that benefited independent developers alongside the platform owner. Jens Foerderer et al. similarly show that Google’s entry into photography apps with Google Photos on Android increased user attention and overall demand for photography apps, followed by greater complementor innovation and entry into adjacent categories.[16] Evidence from video-game console ecosystems points in the same direction: strong first-party titles often expand a platform’s installed base, increasing the potential market for third-party developers even when those developers also compete with first-party games.[17]

More recent experimental evidence from e-commerce reinforces these findings. Chiara Farronato, Andrey Fradkin, and Alexander MacKay conduct a field experiment that hides Amazon-owned private-label brands from shoppers and simulates counterfactual equilibria.[18] In the product categories they study, removing Amazon brands reduces consumer surplus by 5.5% in the short run. Only a small share of that loss reflects higher prices by other sellers; most of the welfare reduction stems from lost variety and diminished consumer valuation of private-label options.[19] Notably, the authors also find that attempts to “correct” potential self-preferencing by demoting private labels in search rankings do not generate consumer-surplus gains.[20]

Taken together, this evidence cautions strongly against treating self-preferencing as presumptively harmful. The empirical record instead shows that welfare effects depend on context and often prove positive—precisely the pattern that supports an effects-based approach, rather than a blanket ban.

C. Scale and Coordination Drive Platform Performance

The MOIT’s focus on platform “operating mechanisms” as a basis for intervention risks confusing sources of consumer value with evidence of competitive harm. Many practices that may appear to be “influence” or “steering” reflect scale economies and coordinated investment—the same efficiencies that vertical integration often delivers. The broader empirical literature on vertical integration consistently shows that integration can reduce transaction costs and improve performance, especially where coordination and quality assurance matter.[21]

These effects carry particular weight in platform markets. Integrated logistics, standardized fulfillment, and unified quality-control systems can produce faster delivery, more reliable service, and lower per-transaction costs at scale. Fragmented providers cannot easily replicate these outcomes without similar coordination.[22] Treating such efficiencies as suspect simply because they disadvantage less efficient rivals risks shifting competition policy toward protecting competitors rather than consumers. A more coherent standard would examine whether a specific practice plausibly raises quality-adjusted prices, reduces output, or forecloses efficient entry. It also would require evidence on those effects before condemning integration or preferential design choices.

IV. Technical Critique of Vietnam’s Proposed Article 27 Amendments

Proposed supplemental Article 27(2) introduces five separate clauses, subparagraphs (a) through (e), to regulate the conduct of dominant digital platforms. Each clause warrants careful scrutiny under the error-cost framework[23] because several risk prohibiting conduct that lacks clear evidence of harm and, in some cases, has demonstrated pro-competitive effects.[24] Where conduct produces ambiguous or context-dependent outcomes, competition law’s case-by-case analysis offers a more reliable and proportionate approach than rigid regulatory mandates.

A. Clause (a): Self-Preferencing Through Ranking, Algorithms, or Technical Specifications

A prohibition on prioritizing a platform’s own products or services through rankings, algorithms, or technical design choices overlooks that product design itself is a central dimension of competition. In digital markets, “self-preferencing” often allows platforms to integrate services in ways that improve performance and usability.[25] E.g., when a search engine displays a map directly in response to a restaurant query, it favors its own mapping service, but it also delivers faster, more useful results than a list of links to third-party sites. Treating such design choices as inherently suspect risks harming consumers by forcing platforms to abandon efficient and value-enhancing product improvements.[26].

The MOIT’s proposal also suggests that self-preferencing through pricing algorithms produces negative competitive effects. The economic literature does not support that conclusion. In a comprehensive review, economists Emilie Feyler and Veronica Postal observe that:

There is no consensus from the economic literature on whether procompetitive benefits or possible anticompetitive considerations prevail in the context of self preferencing algorithms used by digital platforms. Nor is there a consensus on the welfare effects of a policy intervention to correct bias in algorithmic recommendations. Determining the net impact of self-preferencing algorithms on competition and consumer welfare requires individualized analysis accounting for the workings of specific algorithms, competitive context, and market environment.[27]

Vietnam’s proposal further cites “technical specifications” as a potential basis for discriminatory conduct. This approach raises serious risks, as it could compel levels of technical interoperability that undermine system integrity and security.[28] In mobile ecosystems, Apple’s decision to keep iMessage proprietary constitutes a form of self-preferencing, yet it supports a tightly integrated and secure user experience that many consumers deliberately choose over more open alternatives.[29] Mandating access to technical specifications without regard to security, branding, or system design would encourage homogenization and erode the product differentiation that drives competition and innovation.[30]

B. Clause (b): ‘Unreasonable’ Terms and Transaction Conditions

The draft would prohibit platforms from “imposing unreasonable terms and conditions” related to pricing, payment methods, warranties, or other contractual provisions. The concept of “unreasonable” lacks clear economic grounding and introduces significant legal uncertainty for regulated firms. As ICLE has noted in the context of European competition policy, standards based on “fairness” or “reasonableness” resist principled definition and risk functioning as open-ended licenses for discretionary regulatory intervention.[31]

Legal tests built on concepts such as “good faith” or “fair dealing” create persistent uncertainty for market participants. A digital platform that charges a 30% commission to fund app-store security, curation, and infrastructure may view that price as efficient and pro-consumer. A regulator could nonetheless deem it “unreasonable” under an undefined standard. Requiring platforms to defend routine commercial terms against a regulator’s subjective view of fairness would, in effect, transform Vietnam’s competition enforcer into a price-setting authority. That uncertainty would likely deter foreign platforms from introducing new features or business models in Vietnam, given the risk of retrospective findings of “unreasonableness.”

C. Clause (c): Tying and Forced Service Registration

The prohibition on “imposing or forcing users to register, use, or maintain one or more additional services” targets tying and bundling practices.[32] In digital markets, however, firms often compete through bundles, and consumers frequently benefit from integrated offerings that reduce transaction costs and user friction.

The MOIT draft does not distinguish between coercive tying that can foreclose rivals and efficiency-enhancing bundling that benefits consumers. In many settings, integration serves technical and security functions rather than exclusionary aims. For example, allowing third-party applications to run in the background without native mobile operating-system controls can materially reduce battery life and weaken data-privacy protections. Mandating unbundling in such cases would degrade device performance and user experience. ICLE’s research on the U.K. Competition and Markets Authority’s mobile-ecosystem inquiry further indicates that users who remain within bundled ecosystems typically do so because they value the integrated experience, not because platforms lock them in.[33]

D. Clause (d): Multi-Homing and Switching Barriers

The draft seeks to prevent practices that limit business users’ ability to access alternative platforms. Although the MOIT frames this concern in terms of “lock-in,” evidence from global markets shows that multi-homing is the norm rather than the exception. Enterprise customers routinely operate across AWS, Microsoft Azure, and Google Cloud. Developers commonly build for both iOS and Android.[34]

Tools that facilitate data portability and app switching signal active competition, not monopoly power. Firms with durable market power rarely invest in mechanisms that make exit easier; competitors do. Vietnam should therefore approach interoperability mandates with caution. Requirements that push platforms toward a single, homogenized model risk eliminating the diversity of platform approaches—e.g., Apple’s curated ecosystem alongside Google’s more open model—that gives Vietnamese consumers meaningful choice and drives innovation.

E. Clause (e): Mandatory Data Access and Fees

The final proposed clause addresses refusals to provide, or the imposition of allegedly unreasonable conditions or fees for, access to data generated by business users.[35] This provision raises particularly complex technical issues because it intersects directly with data protection, intellectual-property rights, and cybersecurity.

ICLE’s analysis of the European Union’s Digital Markets Act indicates that mandated data access can function as a persistent “live wire” into user accounts.[36] Continuous or real-time access can allow third parties—including potentially malicious actors—to extract communications, media, or location data without further user involvement.[37] Vietnam’s draft does not include adequate safeguards to prevent these risks in the pursuit of increased “contestability.”

Mandated data sharing can also weaken investment incentives. Platforms invest heavily in collecting, cleaning, and structuring data so it can support secure services and advanced analytics.[38] Broad access obligations risk discouraging these investments by forcing firms to share the results of costly data-preparation efforts without clear limits or compensation

V. The Costs and Risks of Importing Ex Ante Digital Regulation

Vietnam’s proposed amendments align with a broader international shift toward ex ante regulation of digital platforms. The EU’s DMA, the UK’s DMCCA, and Germany’s Section 19a regime illustrate this approach, alongside ongoing legislative efforts or debates in jurisdictions such as India, South Korea, Japan, and Brazil.[39]

These frameworks often appear as “best practices” or evidence of global convergence on digital regulation. Closer examination of their implementation, however, reveals significant economic risks, unmet policy objectives, and material geopolitical implications. By adopting this model, Vietnam risks importing an untested and highly contested regulatory framework that may shield less efficient competitors, rather than target demonstrable consumer harm or well-defined market failures.

A. Geopolitical Risks of Targeting Global Digital Platforms

Digital platform regulations, even when not designed to do so, tend to fall disproportionately on U.S.-based companies and therefore carry meaningful geopolitical risk. Critics have characterized the DMA as a politicized instrument—even as a “non-tariff attack,”[40] aimed at constraining U.S. technological leadership.[41] Of the seven gatekeepers designated by the European Commission, five are U.S. companies: Alphabet, Amazon, Apple, Meta, and Microsoft.[42] Regardless of intent, ex ante digital regulation in practice places the greatest compliance burdens on American firms, creating nontrivial geopolitical exposure.

The geopolitical context has shifted sharply following policy changes adopted by the Trump administration in 2025. The U.S. government no longer treats foreign digital regulations as purely domestic policy choices. It now frames them as discriminatory measures—described as “unfair taxes” or “overseas extortion”—directed at U.S. companies. In February 2025, President Donald Trump signed an executive order directing the Office of the U.S. Trade Representative to continue investigations into digital services taxes and to consider responsive measures, including tariffs, against foreign penalties or regulatory actions deemed discriminatory or disproportionate.[43]

Senior U.S. officials have reinforced this stance. Vice President J.D. Vance has publicly warned U.S. allies against expansive regulation of artificial intelligence and digital platforms, explicitly criticizing the DMA, the Digital Services Act (DSA), and the General Data Protection Regulation (GDPR). Speaking at the AI Action Summit in Paris in February 2025, Vance described European-style digital regulation as an innovation deterrent that the United States would not accept.[44]

These geopolitical risks are no longer theoretical. In December 2025, Secretary of State Marco Rubio announced visa restrictions on five European individuals, including a former EU commissioner, whom the U.S. State Department linked to the development and enforcement of the DSA.[45] Rubio described the targeted officials as “agents of the global censorship-industrial complex” who had pressured U.S. platforms and harmed American firms. Whatever the merits of this action, it signals a U.S. willingness to respond not only through trade measures but also by imposing personal consequences on foreign regulators associated with laws perceived as adverse to U.S. interests.

Basic law & economics principles counsel that regulation is justified only when expected benefits exceed expected costs. In the current geopolitical environment—especially given the more confrontational posture of the U.S. administration—these macroeconomic and diplomatic risks warrant careful consideration in any serious cost–benefit assessment of Vietnam’s proposed digital competition amendments.

B. Regulatory Drag and the EU’s Digital Productivity Gap

The Draghi Report observes that “the productivity gap between the U.S. and the EU is largely explained by the tech sector.”[46] The United States has fostered an environment of permissionless innovation, while Europe has layered on dense regulatory constraints.[47] The outcome is visible: Europe has produced no counterparts to Google, Apple, or Amazon.

The Digital Markets Act has also reshaped products in ways that reduce consumer welfare. To comply with the DMA, Google altered the integration of Google Maps into Google Search in the EU. European users now see a static thumbnail with limited functionality rather than a fully interactive map, creating a slower and more fragmented experience that requires additional, unintuitive steps to complete routine tasks.[48] The same negative effect is visible regarding search results for flights and hotels, where additional, counter-intuitive steps were added due to DMA’s prohibition of “self—preferencing.” Similar degradation appears in flight and hotel search results, where the DMA’s restrictions on self-preferencing have reduced integration and usability.[49]

Regulatory uncertainty under the DMA has also delayed the rollout of advanced AI features in Europe and imposed what functions as an “innovation tax” on designated gatekeepers. When firms face unpredictable compliance obligations and potential fines tied to global turnover, they rationally delay, narrow, or geo-fence product launches. This “regulatory chill” lowers the expected value of experimentation and raises the option value of waiting. The practical cost of regulation thus extends beyond compliance spending to include foregone or postponed product improvements for local users.

Recent examples illustrate this pattern. Apple paused the release of Apple Intelligence in the EU amid concerns that DMA interoperability requirements could force design changes that weaken device security.[50] Meta delayed the launch of Threads in the EU for several months, citing uncertainty over DMA limits on combining user data across services such as Instagram and Threads, and redesigned the product to comply.[51] Google has similarly reported that DMA-related reengineering and documentation burdens can delay EU launches—particularly AI-driven search features and integrated modules—by up to a year.[52] These delays reinforce a broader pattern of digital lag in Europe relative to markets such as the United States.

Vietnam’s digital transformation depends on rapid adoption and deployment of advanced technologies developed worldwide. Prescriptive rules that degrade consumer experience, lock business models in place, and deter experimentation would slow that transformation. Countries seeking sustained economic growth should avoid regulatory approaches that substitute rigidity for innovation and evidence-based competition policy.

C. Lessons from the UK’s DMCC

Observers sometimes describe the U.K.’s Digital Markets, Competition and Consumers Act (DMCC) as a more flexible form of ex ante regulation.[53] The U.K.’s experience in 2025 nonetheless highlights the political and institutional instability that discretionary regimes can create.[54]

In response to economic stagnation, the U.K. government applied sustained political pressure on regulators to adopt an explicit “pro-growth” mandate. Prime Minister Keir Starmer publicly emphasized that regulators must place economic growth at the center of their decision-making,[55] and the government required agencies to commit to measurable actions that support business confidence and investment.[56] In parallel, the Competition and Markets Authority (CMA) introduced a package of institutional and procedural reforms—branded as the “4 Ps” of pace, predictability, proportionality, and process. These reforms included changes to merger-review timelines, clearer jurisdictional thresholds, and a new mergers charter designed to align regulatory practice with the government’s growth objectives.[57]

The government’s strategic guidance also stressed accountability, predictability, and collaboration with industry, and urged regulators to exercise new DMCC powers with “particular care,” especially in fast-moving technology markets. This episode shows that even in the United Kingdom—a jurisdiction with relatively strong legal and administrative institutions—ex ante digital regimes remain politically fragile and subject to rapid recalibration.

Vietnam’s institutional environment is less insulated from these pressures. Broad, discretionary platform rules therefore carry a heightened risk of unintended consequences, including innovation drag, rising compliance costs, weaker consumer outcomes, and slower productivity growth. Regulatory modesty—through narrowly tailored, evidence-based intervention—offers a more durable and prudent path than ambitious, early-stage, and untested regulatory frameworks.

VI. The Risks of Imposing a De Facto Duty to Deal in Data

The MOIT proposal would prohibit dominant digital platforms from “abusing business user data,” including by “refusing to provide or imposing unreasonable conditions or fees” on access to data generated through legitimate business activities on the platform. As discussed in Section IV.B, this provision grants Vietnam’s competition authority broad remedial discretion. Enforcement could therefore result in mandatory data sharing or interoperability obligations that risk harming consumers rather than promoting competition.[58]

The proposal also creates a serious risk of regulatory overlap with the Law on Digital Transformation (LODT).[59] The LODT establishes a sector-specific framework for regulating digital platforms, including rules governing data access and portability for designated “dominant platforms.”[60] If MOIT adopts the proposed amendments, platforms could face parallel investigations and sanctions for the same conduct under two separate legal regimes administered by different authorities. As Giuseppe Colangelo has explained in the context of the DMA, overlapping enforcement without clear boundaries breeds fragmentation and legal uncertainty.[61] When a specialized regime already governs conduct such as data access, layering competition-law remedies on top undermines the coherence and effectiveness of that framework.

This duplication creates more than procedural inefficiency. It imposes substantive compliance burdens that can chill investment and innovation. A platform could comply fully with the LODT’s data-access requirements and still face competition-law liability for allegedly “unreasonable” conditions under the Competition Law. That regulatory fog makes compliance unpredictable and deters long-term investment. MOIT should therefore remove these platform-specific data provisions and rely on existing competition-law tools to address demonstrable exclusionary conduct on a case-by-case basis. Policymakers should also subject the LODT’s data-access provisions to a regulatory impact assessment after implementation.[62]

The proposed amendments also risk imposing a de facto duty to deal in data, premised on the assumption that data generated on a platform functions as a public good, rather than as a proprietary asset created through substantial investment. In practice, the value of business-user data often derives from a platform’s aggregation, analytics, and security capabilities. Mandating access to data “generated from business activities” raises fundamental questions about ownership and scope.[63] E.g., when a user enters payment information for an in-app purchase, that data reflects the platform’s payment infrastructure as much as the developer’s activity.

Forcing platforms to share such data for free or under regulated “reasonable” fees would institutionalize free riding. Business users would have incentives to rely on the platform’s infrastructure instead of investing in their own data capabilities.[64] Platforms, in turn, would reduce investment in data collection, cleaning, and security if competitors can immediately appropriate the results. That outcome would undermine innovation and conflict with Vietnam’s broader digital-transformation goals.

The vague prohibition on “unreasonable conditions” also threatens privacy and cybersecurity. Platforms often restrict third-party data access to protect users and preserve system integrity. Broad access mandates—backed by antitrust liability—expand attack surfaces and create new risks. Strong privacy and security protections represent valued product features,[65] not pretexts for exclusion.[66] Forcing platforms to interoperate with numerous third parties that lack robust safeguards increases the likelihood of breaches and system failures.[67]

Recent experience illustrates these risks. In July 2024, a faulty software update from CrowdStrike triggered one of the largest Windows outages on record, disrupting airports, hospitals, and other critical services worldwide.[68] That incident followed longstanding European Commission requirements that Microsoft grant third-party security vendors privileged system access. Competition intervention increased supplier access, but it also amplified systemic risk. By contrast, Apple limits kernel access precisely to preserve security and reliability. Vietnam should approach data-access mandates with similar caution, particularly where consumer trust, privacy, and cybersecurity are at stake.[69]

VII. Conclusion: Innovation-Friendly Competition Policy for Vietnam

Vietnam has strong reasons to foster its digital economy through targeted and agile measures rather than by importing an untested regulatory model whose costs already appear in other jurisdictions. The Law on Digital Transformation and proposed amendments to the Competition Law should place consumer welfare and innovation ahead of concerns about redistributing profits among competitors.

Vietnam should therefore maintain an effects-based competition policy grounded in the following principles:

• Prioritize Consumer Welfare: Distinguish conduct that disadvantages rivals because of superior products from conduct that harms consumers by restricting choice or raising quality-adjusted prices.
• Respect Platform Autonomy: Recognize that firms design their platforms and face the strongest incentives to deliver services that users value.
• Adopt Evidence-Based Standards: Avoid vague concepts such as “unreasonable” or “fair.” Base enforcement on economic analysis, empirical evidence, and industry-specific conditions.
• Favor Light-Touch Remedies: Prefer targeted cease-and-desist orders over prescriptive mandates that force product redesign and risk stifling innovation.
• Protect Privacy and Security: Ensure that competition interventions do not weaken data protection or create new “live wire” vulnerabilities for users.

Regulation should remain a measure of last resort, applied only where markets demonstrably fail. By maintaining a clear, predictable, and proportionate regulatory framework, Vietnam can continue to attract investment and promote dynamic competition that supports long-term growth and digital transformation. This approach reflects not only sound economics but also strategic foresight: innovation flourishes where rules provide clarity and restraint, not where regulation becomes a blunt tool for market engineering.

[1] Ministry of Industry & Trade (Viet.), Policy Dossier for the Draft Law Amending and Supplementing Several Articles of Law No. 23/2018/QH14 of June 12, 2018, on Competition (Competition Law) (Jan. 21, 2026), https://moit.gov.vn/du-thao-van-ban/ho-so-chinh-sach-du-an-luat-sua-du-an-luat-sua-doi-bo-sung-mot-so-dieu-cua-luat-thuong-mai-luat-canh-tranh-luat-quan-ly-.html [hereinafter “Public Consultation”]. Our comments were based primarily on the proposed statutory text available in the Download section, File No. 5, of the above link for the Public Consultation.

[2] Lazar Radic, Comments of the International Center for Law & Economics: Vietnam’s Draft Law on Digital Transformation—A Road to Hell Paved with Good Intentions, Int’l Ctr. for L. & Econ. (Oct. 20, 2025), https://laweconcenter.org/wp-content/uploads/2025/10/Vietnam-open-letter.pdf.

[3] Id.

[4] Public Consultation, supra note 1.

[5] To Lam, Institutional and Legal Breakthroughs Are Needed for the Country to Thrive, Viet Bao (May 4, 2025), https://vietbao.vn/en/tong-bi-thu-to-lam-dot-pha-the-che-phap-luat-de-dat-nuoc-vuon-minh-540884.html.

[6] See generally Nicolas Petit, Big Tech and the Digital Economy: The Moligopoly Scenario (2020).

[7] These concepts originate from European-style regulation (e.g., the Digital Markets Act). See Giuseppe Colangelo, DMA Begins, 11 J. Antitrust Enf’t 116 (2023).

[8] Public Consultation, supra note 1.

[9] Geoffrey A. Manne, Against the Vertical Discrimination Presumption, Concurrences No. 2-2020, art. no. 94267 (May 1, 2020).

[10] Jonathan M. Barnett, The Host’s Dilemma: Strategic Forfeiture in Platform Markets for Informational Goods, 124 Harv. L. Rev. 1861 (2011).

[11] Oliver E. Williamson, Transaction-Cost Economics: The Governance of Contractual Relations, 22 J.L. & Econ. 233 (1979).

[12] See, e.g., Manne, supra note 9.

[13] Brian Albrecht & Geoffrey A. Manne, Self-Preferencing Isn’t a Sin. It’s Often the Way Competition Works., Truth on the Market (Aug. 20, 2025).

[14] Id.

[15] Zhuoxin Li & Ashish Agarwal, Platform Integration and Demand Spillovers in Complementary Markets: Evidence from Facebook’s Integration of Instagram, 63 Mgmt. Sci. 3438 (2017).

[16] Jens Foerderer et al., Does Platform Owner’s Entry Crowd Out Innovation? Evidence from Google Photos, 29 Info. Sys. Res. 444 (2018).

[17] Carmelo Cennamo, Hakan Ozalp & Tobias Kretschmer, Platform Architecture and Quality Trade-offs of Multihoming Complements, 29 Info. Sys. Res. 461 (2018).

[18] Chiara Farronato, Andrey Fradkin & Alexander MacKay, Vertical Integration and Consumer Choice: Evidence from a Field Experiment, Nat’l Bureau of Econ. Research Working Paper No. 34135 (Aug. 2025).

[19] Id.

[20] Id.

[21] Francine Lafontaine & Margaret Slade, Vertical Integration and Firm Boundaries: The Evidence, 45 J. Econ. Literature 629 (2007).

[22] See, e.g., Sam Bowman & Geoffrey A. Manne, Platform Self-Preferencing Can Be Good for Consumers and Even Competitors, Truth on the Market (Mar. 4, 2021).

[23] “The objective of the error-cost framework is to ensure that regulatory rules, enforcement decisions, and judicial outcomes minimize the expected cost of (1) erroneous condemnation and deterrence of beneficial conduct (‘false positives,’ or ‘Type I errors’); (2) erroneous allowance and under-deterrence of harmful conduct (‘false negatives,’ or ‘Type II errors’); and (3) the costs of administering the system (including the cost of making and enforcing rules and judicial decisions, the costs of obtaining and evaluating information and evidence relevant to decision-making, and the costs of compliance).” Geoffrey A. Manne, Error Costs in Digital Markets, in The GAI Report on the Digital Economy 34 (2020).

[24] See, infra, Section III.B. Most practices the proposal would ban are vertical restraints—agreements or other constraints between firms at different levels of the production chain—and therefore warrant analysis under a rule-of-reason framework. See Jonathan Barnett, Does the European Union’s Digital Markets Act Provide an Appropriate Model for Maintaining Competition in California’s Innovation Economy? 15 (report commissioned by the Chamber of Progress, Dec. 2023), http://www.clrc.ca.gov/pub/2024/MM24-05.pdf.

[25] Manne, supra note 23, at 38-39.

[26] See, infra, Section VI regarding Google’s downstream effects on users following implementation of the DMA’s prohibition on self-preferencing.

[27] Emilie Feyler & Veronica Postal, Can Self-Preferencing Algorithms Be Pro-Competitive?, CPI Antitrust Chron. 5 (June 2023), https://www.competitionpolicyinternational.com/wp-content/uploads/2023/06/5-can-selfpreferencing-algorithms-be-pro-competitive-emilie-feyler-veronica-postal.pdf.

[28] Miko?aj Barczentewicz, The Digital Markets Act Shouldn’t Mandate Radical Interoperability, Truth on the Market (May 19, 2021), https://truthonthemarket.com/2021/05/19/the-digital-markets-act-shouldnt-mandate-radical-interoperability.

[29] Geoffrey A. Manne, Dirk Auer & Mário A. Zuñiga, Comments of the International Center for Law & Economics on CMA’s Proposal to Designate Apple and Google with Strategic Market Status, Int’l Ctr. for L. & Econ. (Aug. 20, 2025), https://laweconcenter.org/resources/icle-comments-to-uk-cma-on-sms-designations-for-mobile-ecosystems.

[30] Id.

[31] Giuseppe Colangelo, Fairness and Ambiguity in EU Competition Policy, Int’l Ctr. for L. & Econ. (ICLE White Paper No. 2023-02-15), https://laweconcenter.org/resources/fairness-and-ambiguity-in-eu-competition-policy.

[32] Public Consultation, supra note 1.

[33] Geoffrey A. Manne, Dirk Auer & Mário A. Zuñiga, Comments of the International Center for Law & Economics on CMA’s Proposal to Designate Apple and Google with Strategic Market Status, Int’l Ctr. for L. & Econ. (Aug. 20, 2025), https://laweconcenter.org/resources/icle-comments-to-uk-cma-on-sms-designations-for-mobile-ecosystems.

[34] Sami Hyrynsalmi, Arho Suominen & Matti Mäntymäki, The Influence of Developer Multi-Homing on Competition Between Software Ecosystems, 111 J. Syst. Softw. 119 (2016).

[35] Public Consultation, supra note 1.

[36] Miko?aj Barczentewicz, ICLE Comments on the Interplay Between DMA and GDPR, Int’l Ctr. for L. & Econ. (Dec. 4, 2025), https://laweconcenter.org/resources/icle-comments-on-the-interplay-between-dma-and-gdpr.

[37] Id.

[38] Nathalie Jorzik, Paula J. Kirchhof & Frank Mueller-Langer, Industrial Data Sharing and Data Readiness: A Law and Economics Perspective, 57 Eur. J. L. & Econ. 181 (2024).

[39] Geoffrey A. Manne et al., ICLE Comments to the U.K. Competition & Markets Authority on SMS Designations for Mobile Ecosystems, Int’l Ctr. for L. & Econ. (Aug. 20, 2025), https://laweconcenter.org/resources/icle-comments-to-uk-cma-on-sms-designations-for-mobile-ecosystems.

[40] Robert D. Atkinson, Letter to the Trump Administration Regarding Non-Tariff Attacks on U.S. Tech Firms and Industries, Info. Tech. & Innovation Found. (July 2, 2025), https://itif.org/publications/2025/07/02/letter-regarding-non-tariff-attacks-on-ustech-firms-and-industries.

^[41] Miko?aj Barczentewicz, The Digital Markets Act as an EU Digital Tax: When Compliance Costs Dwarf Regulatory Estimates, Truth on the Market (July 8, 2025), https://laweconcenter.org/resources/the-digital-markets-act-as-an-eu-digital-tax-whencompliance-costs-dwarf-regulatory-estimates.

^[42] European Comm’n, Gatekeepers, https://digital-markets-act.ec.europa.eu/gatekeepers_en (last visited Jan. 27, 2026).

^[43] The White House, Defending American Companies and Innovators from Overseas Extortion and Unfair Fines and Penalties, Presidential Actions (Feb. 21, 2025), https://www.whitehouse.gov/presidential-actions/2025/02/defending-american-companies-and-innovators-from-overseas-extortion-and-unfair-fines-and-penalties.

[44] Siladitya Ray, JD Vance Knocks EU’s Regulation of U.S. Tech Giants: “America Cannot Accept That”, Forbes (Feb. 11, 2025), https://www.forbes.com/sites/siladityaray/2025/02/11/jd-vance-knocks-eus-regulation-of-us-tech-giants-america-cannot-accept-that.

^[45] Kim Marcrael, U.S. Sanctions Former EU Official Over Digital-Content Law, Wall St. J. (Dec. 24, 2025), https://www.wsj.com/world/europe/u-s-sanctions-former-eu-official-over-digital-content-law-c41f574c.

^[46] Mario Draghi, The Future of European Competitiveness – Part A: A Competitiveness Strategy for Europe 5 (European Comm’n Sept. 2024), https://commission.europa.eu/document/download/97e481fd-2dc3-412d-be4c-f152a8232961_en?filename=The%20future%20of%20European%20competitiveness%20_%20A%20competitiveness%20strategy%20for%20Europe.pdf.

^[47] Id. at 6 (“The problem is not that Europe lacks ideas or ambition. We have many talented researchers and entrepreneurs filing patents. But innovation is blocked at the next stage: we are failing to translate innovation into commercialisation, and innovative companies that want to scale up in Europe are hindered at every stage by inconsistent and restrictive regulations”).

[48] Impact of the Digital Markets Act (DMA) on Consumers Across the European Union: Results from a Survey with 5,000 Consumers, Nextrade Grp. (Sept. 2025), https://www.nextradegroupllc.com/impact-of-the-dma-on-eu-consumers.

[49] Id.

[50] Akshaya Asokan, Apple to Delay AI Rollout in Europe, BankInfoSecurity (June 21, 2024).

[51] Egl? Markevi?i?t?, Consumer Waiting Game: Why Do Tech Products Launch Later in Europe?, Euronews (Sept. 26, 2025), https://www.euronews.com/next/2025/09/26/consumer-waiting-game-why-do-tech-products-launch-later-in-europe.

[52] Cynthia Kroet, Google’s AI Feature on Hold in Most EU Member States Due to “Strict Rules”, Euronews (Apr. 1, 2025), https://www.euronews.com/next/2025/04/01/googles-ai-feature-on-hold-in-most-eu-member-states-due-to-strict-rules.

^[53] Mario Zúñiga, Parsing Brazil’s ‘More Flexible’ Approach to Digital Markets, Truth on the Market (Feb. 5, 2025), https://truthonthemarket.com/2025/02/05/parsing-brazils-more-flexible-approach-to-digital-markets.

^[54] Dario Oliveira Neto, Lessons from the UK for Brazil’s Digital Market Strategy, Truth on the Market (July 22, 2025), https://truthonthemarket.com/2025/07/22/lessons-from-the-uk-for-brazils-digital-market-strategy.

^[55] Alistair Smout, UK Pledges Regulatory Overhaul to Try to Win Over Investors, Reuters (Oct. 14, 2024), https://www.reuters.com/world/uk/uk-promises-regulation-overhaul-bid-court-wary-investors-2024-10-13.

^[56] HM Treasury, New Approach to Ensure Regulators and Regulation Support Growth (Oct. 22, 2025), https://www.gov.uk/government/publications/a-new-approach-to-ensure-regulators-and-regulation-support-growth/new-approach-to-ensure-regulators-and-regulation-support-growth-html.

^[57] Id.

^[58] See Competition Law ch. IX, art. 110(4).

^[59] Law No. 148/2025/QH15 (Dec. 11, 2025) (Viet.) (effective July 1, 2026).

^[60] See Huu Tuan Nguyen & Alex Do, Vietnam’s Draft Digital Transformation Law Proposes “Far-Reaching” Paradigm for Digital Platforms, IAPP (Sept. 17, 2025), https://connectontech.bakermckenzie.com/vietnam-vietnams-draft-digital-transformation-law-proposes-far-reaching-paradigm-for-digital-platforms; see also Lazar Radic, ICLE Comments on Vietnam’s Law on Digital Transformation, Int’l Ctr. for L. & Econ. (Oct. 20, 2025), https://laweconcenter.org/resources/icle-comments-on-vietnams-digital-transformation-bill.

^[61] Giuseppe Colangelo, The Digital Markets Act and EU Antitrust Enforcement: Double & Triple Jeopardy, ICLE White Paper (Mar 23, 2022), https://laweconcenter.org/resources/the-digital-markets-act-and-eu-antitrust-enforcement-double-triple-jeopardy

[62] Ongoing monitoring and evaluation are core regulatory best practices. They improve regulatory quality and provide an essential check on the exercise of regulatory power. See OECD, Regulatory Impact Assessment 29 (2020), https://www.oecd.org/content/dam/oecd/en/publications/reports/2020/02/regulatory-impact-assessment_0bf78a03/7a9638cb-en.pdf.

[63] See Geoffrey A. Manne & Dirk Auer, Antitrust Dystopia and Antitrust Nostalgia: Alarmist Theories of Harm in Digital Markets and Their Origins, 28 Geo. Mason L. Rev. 1279, 1351 (2021). (“The challenge for firms in data-reliant industries is multidimensional. Not only must they acquire data (and this is not merely a matter of ‘data network effects’), but just as importantly they must also develop the expertise to analyze this data, draw useful insights from it, and turn these insights into successful products. In doing so, acquiring the right data and getting the best out of a firm’s engineers is at least as important as controlling a large amount of data or engineering expertise… Under this light, the resounding success of certain technology platforms appears to be down to their respective ‘dynamic capabilities’ rather than the operation of positive feedback loops.”)

[64] Brian Albrecht & Dirk Auer, Free Riding in Mobile Ecosystems, Int’l Ctr. for L. & Econ. (Dec. 2, 2025), https://laweconcenter.org/resources/free-riding-in-mobile-ecosystems.

[65] In Epic Games, Inc. v. Apple Inc., the 9th U.S. Circuit Court of Appeals recognized that security and privacy play a decisive role in consumer choice. The court noted that 50% to 62% of iPhone users and 76% to 89% of iPad users worldwide consider security and privacy important when purchasing a device. Even Epic’s CEO testified that he chose an iPhone over an Android device in part because it offers “better security and privacy.” The district court further found that Apple’s creation of a “trusted app environment” leads users to make greater use of their devices. See Epic Games, Inc. v. Apple Inc., No. 25-2935, 53 (9th Cir. 2025).

[66] See Margrethe Vestager, A Whack-a-Mole Approach to Big Tech Won’t Do, Says Europe’s Antitrust Chief, Economist (June 4, 2024), https://www.economist.com/by-invitation/2024/06/04/a-whack-a-mole-approach-to-big-tech-wont-do-says-europes-antitrust-chief (arguing that “asking platforms to open up their ecosystem, for instance, does not mean they have to compromise the security of their service”).

[67] See Miko?aj Barczentewicz, Does the DMA Let Gatekeepers Protect Data Privacy and Security?, Truth on the Market (Apr. 4, 2024), https://truthonthemarket.com/2024/04/04/does-the-dma-let-gatekeepers-protect-data-privacy-and-security; Mario Zúñiga, The EU Is Determined to Tear Down Apple’s ‘Walled Garden’, Truth on the Market (May 6, 2025), https://truthonthemarket.com/2025/05/06/the-eu-is-determined-to-tear-down-apples-walled-garden.

[68] Bobby Allyn, Brian Mann, Bill Chappell & Fatima Al-Kassab, What We Know About the Computer Update Glitch Disrupting Systems Around the World, Nat’l Pub. Radio (July 19, 2024), https://www.npr.org/2024/07/19/g-s1-12222/microsoft-outage-banks-airlines-broadcasters.

[69] Jowi Morales, Microsoft’s EU Agreement Means It Will Be Hard to Avoid CrowdStrike-Like Calamities in the Future, Tom’s Hardware (July 22, 2024), https://www.tomshardware.com/software/windows/microsofts-eu-agreement-means-it-will-be-hard-to-avoid-crowdstrike-like-calamities-in-the-future.

ICLE Senior Scholar Selcukhan Unekbas appeared on the Shaping Competition in the Digital Age (SCiDA) podcast with Pankhudi Khandelwal to discuss their research at the European University Institute. Unekbas presented empirical work finding little evidence for “killer acquisitions” and outlined a proposed “synamically as-efficient competitor” test for EU merger law, while the episode also covered interoperability lessons for digital regulation, duties-to-deal doctrine under Article 102, and possible remedies following the European Commission’s Google Adtech decision. Audio of the full episode is embedded below.

Executive Summary

America’s telephone network was built on copper wire. For much of the 20th century, that infrastructure carried nearly every call. Today, consumers rely on mobile service, Voice over Internet Protocol (VoIP), and fiber broadband. Yet federal and state rules still require carriers to maintain aging copper networks—even where few customers remain. The evidence shows these requirements impose large and growing costs with shrinking benefits, and reform is economically justified.

The consumer shift is clear. From 2014 to 2024, subscribers using copper last-mile connections fell 81%, from nearly 66 million to about 12.5 million. Over the same period, mobile subscriptions rose from 322 million to 391 million. By 2024, 79% of U.S. adults lived in wireless-only households, while fewer than 1% relied exclusively on landlines. Consumer behavior has already determined the outcome; regulation has not caught up.

Maintaining copper networks for this shrinking user base is increasingly expensive. In 2024, AT&T reported spending roughly $6 billion annually—about 5% of total revenue—to keep its copper network operating. These costs include maintenance, energy use, cooling, and real estate for oversized wire centers. In California alone, AT&T spent more than $1 billion in 2023 maintaining copper infrastructure serving fewer than 5% of households, now closer to 3%. By contrast, Verizon’s migration of 4.5 million circuits to fiber produced about $180 million in annual operating savings and a 60% reduction in maintenance dispatches. Similar trends appear across mid-sized and regional carriers.

The disparity reflects copper’s physical limitations. Copper corrodes, absorbs moisture, and degrades under environmental stress. Failures require costly maintenance dispatches—“truck rolls”—that often cost hundreds of dollars per visit. Fiber transmits light rather than electricity, making it more reliable and cheaper to operate. Industry analysis estimates that all-fiber networks cost about $91 less per home annually than copper-based DSL networks, and AT&T reports fiber costs roughly 35% less per subscriber to maintain.

Energy consumption widens the gap. Copper systems require continuous power and cooling for central-office equipment. AT&T estimates its copper-to-fiber transition saved about 340,000 megawatt-hours of electricity in 2024 alone. Altafiber reports copper service consumes 172 kWh annually per subscriber compared with just 6 kWh for fiber—a 97% reduction. Transitioning the remaining copper subscribers would save an estimated $398 million to $830 million in energy costs alone, while also reducing emissions.

Copper’s commodity value now creates public-safety risks. Prices rose from $2.29 per pound in 2020 to nearly $6 by early 2026. In 2025, AT&T reported nearly 8,700 theft incidents, costing about $76 million in repairs. Industrywide, more than 15,500 theft and sabotage incidents between mid-2024 and mid-2025 disrupted service to over 9.5 million customers, at times affecting 911 systems, hospitals, and military facilities. Estimated repair costs approach $136 million annually, and the social costs are higher.

The infrastructure itself is aging out. Core switching systems were last manufactured decades ago, and replacement parts now come largely from secondary markets. Each year, maintenance becomes harder, more expensive, and less reliable.

The policy question is not whether transition entails costs—it does. The question is whether those costs justify spending billions each year to preserve infrastructure serving a shrinking and voluntarily departing customer base. Removing regulatory barriers to copper retirement would reduce deadweight loss, reallocate capital and labor to next-generation networks, improve resilience and energy efficiency, and align telecommunications policy with today’s competitive marketplace.

I. Introduction and Overview

In 2025, the International Center for Law & Economics (ICLE) filed comments with the Federal Communications Commission (FCC) on copper retirement.[1] Those comments argued that existing FCC rules make it unnecessarily costly to migrate from aging copper networks to next-generation IP-based infrastructure. Current network-change and service-discontinuance regulations impose substantial transaction costs and create deadweight loss by inflating the price of retiring inefficient copper facilities. Removing those impediments would improve consumer welfare and speed infrastructure upgrades.

ICLE therefore urged the FCC to use its Section 10 forbearance authority to waive the network-change notice requirements of Section 251(c)(5) and the service-discontinuance requirements of Section 214 where competitive alternatives exist. Doing so would better reflect modern competition, reduce unnecessary compliance costs, and free private capital for faster deployment of advanced networks.

The ICLE comments explained:

As markets have moved toward mobile and VoIP, maintaining legacy copper lines has become increasingly expensive. With fewer subscribers, the per-subscriber cost for copper lines has increased. Prices have also gone up for standalone voice service over copper lines, leading to even more switching by consumers. This negative feedback loop is unsustainable over the long term, which is why many providers are looking to retire their copper networks.[2]

The filing identified several drivers of rising maintenance costs:

• Operating wire centers to serve a shrinking customer base;
• Growing copper theft, partly driven by higher commodity prices; and
• The declining availability of hardware and cables no longer manufactured or supported.

This issue brief builds on those observations. Evidence from carrier financial disclosures, FCC reports, industry studies, and peer-reviewed research shows three consistent trends: maintenance costs are large and rising, consumers have shifted decisively to alternative technologies, and regulatory delay magnifies costs as the subscriber base shrinks and infrastructure deteriorates. The data point to a straightforward conclusion—the commission’s proposed reforms are economically justified.

II. Copper Is No Longer the Consumer Default

The Federal Communications Commission’s most recent “Voice Telephone Services” report captures the scale of the market shift. From 2014 to 2024, subscribers using switched access lines fell 77%, from nearly 73 million to just over 16 million.[3] Subscribers served by copper local-loop—or “last-mile”—connections dropped even faster, declining 81%, from almost 66 million to about 12.5 million.

Figure 1: Number of Subscribers to Voice Telephone Services by Type (000s)

SOURCE: Federal Communications Commission

Copper’s share of the remaining wireline market also collapsed. In 2014, 60% of wireline subscribers relied on copper last-mile connections. By 2024, that figure had fallen to 28%. Over the same period, mobile subscriptions rose from 322 million to 391 million, a 21% increase (see Figure 1).

The Centers for Disease Control and Prevention’s National Health Interview Survey confirms the consumer response. In 2024, roughly 79% of adults lived in wireless-only households, while only 0.9% lived in landline-only households.[4]

III. Maintaining Copper Networks Is Economically Unsustainable

For U.S. telecommunications carriers, maintaining legacy copper networks has shifted from routine upkeep to a major operational burden. Copper once formed the backbone of the nation’s communications system. The rapid spread of fiber and other modern technologies now exposes how costly and inefficient those legacy facilities have become.

Across the industry—large incumbents and regional providers alike—the pattern is consistent. Carriers spend billions each year to operate infrastructure serving a rapidly shrinking customer base. Much of the expense does not vary with usage: utilities must still power equipment, maintain wire centers, dispatch technicians, and hold real estate, even as subscribers leave. As lines disappear, the per-customer cost rises.

The transition to fiber reverses those economics. Companies that migrate customers off copper report fewer repair dispatches, reduced building footprints, and lower operating costs. They also report substantial reductions in electricity use and emissions. Together, the operational, financial, and energy data point in the same direction: copper networks impose escalating costs, while modern networks generate measurable savings and reliability gains.

A. AT&T Spends Billions to Maintain a Shrinking Copper Network

At its 2024 Analyst & Investor Day, AT&T reported spending roughly $6 billion each year in direct operating costs to keep its copper network running—about 5% of the company’s $122 billion in annual revenue.[5] The figure excludes capital expenditures and legacy IT-system costs. About 40% reflects customer-facing expenses, such as installation and service calls, while the remaining 60% consists of fixed geographic costs, including power, maintenance, and real estate.

The same pattern appears at the state level. In California, AT&T spent more than $1 billion in 2023 maintaining its legacy copper network even though fewer than 5% of households in its service territory still relied on traditional copper voice service.[6] Today, that infrastructure serves only about 3% of households.[7]

Subscriber counts continue to fall rapidly. As of year-end 2024, AT&T had 3.3 million network-access lines and 127,000 DSL subscribers remaining, down from 4.2 million lines and 210,000 DSL subscribers the year before—a decline of 983,000 lines in a single year.[8]

In January 2025, CEO John Stankey announced the company would seek FCC approval to stop selling legacy products at roughly 1,300 of its approximately 4,600 wire centers.[9] The FCC has since approved retirement of copper facilities across more than 30% of AT&T’s footprint (excluding California), with discontinuations targeted to begin in late 2026 and full retirement planned by the end of 2029.[10]

B. Verizon’s Fiber Shift Slashes Operating and Real-Estate Costs

Verizon has migrated 4.5 million circuits from copper to fiber and fully retired 36 central offices. The company reports roughly $180 million in annual operating savings, driven by fewer maintenance dispatches and lower energy use.[11] Fiber conversion alone reduced maintenance dispatches by about 60%.[12]

Regulatory filings show the same effect. In a 2017 FCC filing seeking approval to retire copper in eight Northeast markets, Verizon documented 3.4 million fewer repair and troubleshooting dispatches between 2012 and 2016 than would have occurred had those customers remained on copper.[13]

The savings extend beyond field operations. Verizon maintains about 50 million square feet of central-office real estate but estimates that 60% to 80% becomes unnecessary in an all-fiber network. Decommissioning legacy Class 5 switches and associated copper infrastructure allows the company to shrink facilities dramatically—reducing buildings that once required up to 13 floors of equipment to just one or two.[14]

C. Mid-Sized and Regional Carriers Are Also Moving from Copper to Fiber

The shift away from copper extends well beyond the largest carriers.

Lumen Technologies lost 310,000 legacy (“other”) broadband subscribers between Q4 2024 and Q4 2025—a 21% decline—leaving 1.16 million total broadband subscribers.[15] Lumen sold its consumer fiber business to AT&T for $5.75 billion, while retaining its copper broadband and voice operations.[16]

Frontier Communications reported a 234,000 drop in copper subscribers during 2024, leaving just 702,000 at year-end.[17] At the same time, fiber broadband revenue rose 25% year over year, including a record 133,000 fiber net additions in Q3 2025.[18] Verizon completed its $20 billion acquisition of Frontier in January 2026, further consolidating the industry’s fiber-first trajectory.[19]

Altafiber experienced similar trends. Its legacy subscriber base fell from 227,000 in 2024 to 187,000 in 2025, a 17% decline.[20]

Newer market entrants show the same direction of change. Brightspeed—formed after acquiring Lumen’s incumbent local-exchange carrier (ILEC) assets across 20 states for $7.5 billion—serves more than 7.3 million homes and businesses and had passed more than 2 million locations with fiber as of April 2025.[21] The company uses what it calls a “reactive” retirement approach: customers experiencing repeated copper failures are migrated to fiber or wireless service.[22]

Consolidated Communications (Fidium) has likewise sought to discontinue legacy voice service. It petitioned the FCC to retire copper voice service at more than 45,000 locations across Maine, New Hampshire, and Vermont, followed by a second petition covering another 61,000 locations in August 2025.[23] Searchlight Capital Partners acquired the company for $3.1 billion in December 2024, and nearly 60% of its footprint has now been upgraded to fiber.[24]

D. Fiber Networks Dramatically Reduce Energy Use and Emissions

Fiber-optic networks deliver large gains in energy efficiency and environmental performance relative to legacy copper and coaxial systems. Copper-based POTS (“plain old telephone service”) networks require substantial electricity to maintain electrical signals over long distances and to cool heat-intensive equipment—even when traffic volumes are low. Fiber, by contrast, transmits information using light pulses that experience minimal signal loss and generate far less heat, making networks cheaper to operate.

Multiple studies estimate that replacing copper or cable with fiber and wireless can cut carbon emissions by as much as 90% or more. Transitioning the remaining 12.5 million copper subscribers to modern alternatives would produce estimated energy-cost savings of $398 million to $830 million.[25] As data demand continues to rise, retiring these energy-intensive legacy systems offers both economic and environmental benefits, alongside a technological upgrade.

The energy burden of copper networks is substantial. AT&T’s copper-to-fiber transition saved about 340,000 MWh in 2024—roughly 346 kWh per switched subscriber, or about $66 in energy savings per subscriber.[26] The company projects cumulative savings of 1.06 million MWh from 2024 through 2028, reducing emissions by roughly 740,000 metric tons of CO2 equivalent.[27] AT&T also reports a 70% reduction in energy consumption when neighborhoods move from copper DSL to 1 Gbps fiber service, reflecting both lower power requirements and reduced maintenance needs.[28]

Altafiber reports similar results. Its legacy copper network consumes about 172 kWh annually per subscriber, compared with just 6 kWh for a fiber connection—a 97% reduction.[29] At current electricity prices, that equals roughly $32 in annual energy savings per subscriber.[30] Operating the copper network accounts for 65% of Altafiber’s total carbon emissions: 48% from powering equipment and 17% from cooling.[31]

Industrywide evidence points in the same direction. Verizon’s 2023 ESG report states that fiber-delivered broadband is at least 100 times more energy-efficient on a kWh-per-gigabyte basis than copper-delivered service.[32] A 2023 lifecycle assessment by Corning found optical fiber’s embodied carbon up to five times lower than a copper wire pair.[33] A 2022 Telefónica lifecycle assessment likewise found the environmental impact per petabyte of fiber traffic 18 times lower than copper—a 94% reduction.[34]

IV. Copper Networks Are Less Reliable and More Expensive to Maintain

Copper networks are inherently more expensive to maintain than fiber because of the medium’s physical limitations. Much of the copper plant installed between the late 19th and mid-20th centuries is vulnerable to moisture, corrosion, and electromagnetic interference.[35] Water intrusion can disable a line for days or weeks, leaving remaining subscribers without service.[36] Fiber transmits light rather than electricity, making it immune to electromagnetic interference and far less susceptible to lightning, power surges, and environmental degradation.[37] As a result, fiber-based—including wireless—networks are more resilient and faster to repair.

Carriers’ own data reflect these differences. AT&T reported at its December 2024 Analyst Day that fiber costs about 35% less per subscriber to maintain than copper because failures occur less often.[38] A 2020 Fiber Broadband Association analysis similarly found annual operating costs of $53 per home passed for all-fiber networks, compared with $144 for DSL—a $91 yearly savings, or roughly $910 over a decade.[39] DSL operators also face churn-related costs about 2.5 times higher than fiber providers.[40]

Copper’s higher failure rate drives a large number of maintenance dispatches, commonly called “truck rolls.”[41] Industry estimates place the cost of a single dispatch at $150 to $500, including fuel, labor, and vehicle depreciation.[42] Reducing these visits frees capital for revenue-producing investment, such as new fiber deployment, rather than ongoing repair work associated with legacy facilities.

Environmental performance follows the same pattern. A January 2025 Ramboll report found that installing fiber infrastructure produces “clear environmental benefits in the long term” relative to maintaining copper and that fiber broadband is at least 100 times more energy-efficient than copper-based service during operation.[43]

V. Copper Theft Imposes Growing Economic and Public-Safety Costs

The copper facilities carriers struggle to maintain have also become valuable targets for theft, especially as commodity prices rise. What once counted as a maintenance problem now creates a security problem.

The consequences extend well beyond carrier balance sheets. Theft has disrupted 911 systems, hospitals, military installations, and local government communications, leaving millions of customers without service and imposing broader social costs.

Regulatory delay compounds the risk. By slowing retirement of legacy copper networks, existing rules leave vulnerable infrastructure in place longer, increasing both economic losses and public-safety exposure.

A. The Scale of Copper Theft Is Driving Costs and Outages

Rising copper prices have made telecommunications infrastructure an increasingly attractive theft target. After the pandemic-era recession in 2020, copper prices climbed from $2.29 per pound to $5.89 per pound by January 2026—an annualized growth rate of roughly 18%.[44]

The operational impact is substantial. AT&T reported nearly 8,700 copper-theft incidents nationwide in 2025, causing about $76 million in repair costs—roughly $8,735 per incident.[45] In California alone, AT&T recorded 2,200 thefts in 2024, up from just 71 in 2021, a thirtyfold increase.[46]

Industrywide data show the same trend. Providers documented 15,540 theft and sabotage incidents between June 2024 and June 2025, disrupting service for more than 9.5 million customers.[47] Using AT&T’s average repair cost, that equals approximately $136 million in annual repair expenses. The pace also accelerated: 9,770 incidents occurred in the first half of 2025, nearly double the 5,770 reported in the second half of 2024.[48] California and Texas accounted for roughly half of all incidents.[49]

The effects extend beyond telecommunications networks. In Los Angeles, the Bureau of Street Lighting reported a tenfold increase in theft-related outages—from 607 incidents in fiscal year 2017–2018 to 6,344 in fiscal year 2021–2022.[50]

B. Copper Theft Now Threatens Critical Infrastructure

Federal authorities have long treated copper theft as more than a property crime. The Federal Bureau of Investigation identified it as a threat to critical infrastructure as early as 2008.[51] Documented incidents have disrupted 911 dispatch systems, law enforcement communications, hospitals, military installations, and schools.[52]

Recent examples underscore the risk. In Contra Costa County, California, copper theft disrupted 911 service for about a week in June 2024.[53] In June 2025, thieves searching for copper damaged Charter’s fiber network in Los Angeles, causing an outage affecting a U.S. military base, emergency dispatch services, police and fire departments, and more than 50,000 people.[54]

States have begun responding. In 2025, 23 states considered infrastructure-protection legislation, and 13 enacted laws increasing penalties for theft and vandalism targeting communications networks.[55]

Economist Edward J. Lopez estimates the broader social harm using a willingness-to-pay framework. Incorporating loss aversion—the higher cost people place on losing service than on gaining it—and network effects affecting those who cannot reach disconnected users, he finds the 5,770 incidents reported in the second half of 2024 alone imposed societal costs between $38 million and $188 million.[56]

VI. Copper Networks Depend on Obsolete Equipment

The copper public-switched telephone network relies on central-office switches that manufacturers no longer produce. The Lucent 5ESS—once one of the most widely deployed Class 5 switches in the United States—entered service in 1982 and was last manufactured in 2003.[57] Nortel, maker of the DMS-100 digital switch introduced in 1979,[58] filed for Chapter 11 bankruptcy in the U.S. Bankruptcy Court for the District of Delaware in 2009.[59] Siemens’ EWSD switching system, first released in 1975, reached end-of-manufacture when Siemens set a final order date of July 15, 2007, for the EWSD core.[60] Support later transferred to Nokia Siemens Networks (now Nokia), but the system is now fully discontinued.[61]

With original manufacturers gone, operators must keep copper networks running using secondary-market vendors and refurbished parts.[62] In one representative case, Tinker Air Force Base maintained a 5ESS switch for years by purchasing replacement components on eBay, because commercial parts had been unavailable for nearly a decade before the system was finally decommissioned in October 2024.[63]

The obsolescence problem extends beyond switching equipment. The SONET/SDH transport systems that carry traffic across copper networks are also disappearing. Major vendors—including Cisco, Ciena, Nokia, and Ericsson—have ended or are phasing out support for legacy SONET/SDH time-division multiplexing transport products, effectively forcing upgrades.[64] Cisco, for example, announced end-of-sale dates for components of its ONS 15454 platform with last support ending in 2019.[65] Analysts likewise reported a 30% drop in carrier spending on SONET/SDH equipment, concluding the technology had been effectively displaced.[66]

Each year, the practical consequences intensify: replacement components grow scarcer, maintenance becomes more expensive, and no new production exists to replenish the supply.

VII. The Policy Case for Allowing Copper Retirement

The evidence presented in this brief points in one direction: requiring continued maintenance of legacy copper networks imposes large costs while delivering shrinking benefits. AT&T alone spends about $6 billion each year—roughly 5% of its revenue—to operate a network serving a steadily declining share of customers.[67] Across the industry, carriers incur additional billions in maintenance expenses, energy consumption, theft losses, and delayed investment in modern infrastructure.

The potential gains from transition are substantial. The Brattle Group estimated in 2024 that completing nationwide fiber deployment would generate $3.24 trillion in net present value and support roughly 380,000 jobs.[68] This brief does not attribute that full amount to the FCC’s proposed reforms. The estimate reflects the value of a complete national fiber buildout. The reforms would instead remove procedural barriers that delay retirement of copper facilities where alternatives exist, redirect billions in annual maintenance spending toward fiber and other modern technologies, and reduce regulatory drag that discourages deployment by incumbent providers. The Fiber Broadband Association similarly estimates all-fiber networks save $91 per home passed each year relative to DSL—across tens of millions of homes, a significant resource misallocation.[69]

International experience reinforces the point. A 2020 WIK-Consult study for the FTTH Council Europe found that lengthy regulatory notice periods—up to five years in some countries—significantly delayed copper switch-off even where fiber was available, recommending shorter timelines once alternatives exist.[70] An Accenture analysis commissioned by Australia’s government-owned broadband operator projected AU$10.4 billion in cumulative GDP growth from 2026 to 2034 from upgrading remaining fiber-to-the-node networks to full fiber.[71] Different institutional structures notwithstanding, the common lesson holds: regulatory certainty accelerates investment.

Transition policy must also account for remaining copper subscribers, who skew older and more rural than the general population.[72] The FCC’s proposed framework addresses this concern through notice requirements and the availability of alternative voice services, including wireless and VoIP. The policy question is not whether transition costs exist—they do—but whether those costs justify continued spending of billions annually to maintain infrastructure serving a shrinking and voluntarily departing customer base.

Resources dedicated to copper maintenance—skilled labor, contractor capacity, and capital—directly compete with fiber deployment and broadband expansion. The proposed reforms would free substantial funding and workforce capacity for investment in next-generation networks with useful lives measured in decades.

VIII. Conclusion

The record shows a widening mismatch between regulation and reality. Consumers have moved to mobile, VoIP, and fiber-based services, yet federal and state rules still require carriers to devote billions each year to infrastructure serving a shrinking and voluntarily departing customer base. Those expenditures divert capital and skilled labor away from fiber and fixed-wireless deployment, reduce network resilience, and prolong exposure to outages, theft, and equipment failure. Each year of delay compounds the problem, as copper facilities deteriorate, parts grow scarce, and maintenance costs rise.

The FCC’s proposed reforms recognize that telecommunications policy must track current technology and competition. Where adequate alternatives exist, preserving legacy barriers to copper retirement neither protects consumers nor promotes efficient investment. Modernizing network-change and service-discontinuance rules would reduce deadweight loss, redirect resources toward durable infrastructure, and speed the transition to more reliable and energy-efficient networks.

The question is no longer whether the transition will occur. It is whether regulation will permit it to proceed in an orderly and economically rational way. The evidence supports reform.

[1] Kristian Stout, Ben Sperry & Eric Fruits, Comments of the International Center for Law & Economics, Reducing Barriers to Network Improvements and Service Changes, Docket No. 25-209; Accelerating Network Modernization, Docket No. 25-208 (Aug. 22, 2025), https://www.fcc.gov/ecfs/document/10822002748400/1.

[2] Id. at 5.

[3] FCC, Voice Telephone Services: Status as of December 31, 2024 (Feb. 2026), https://docs.fcc.gov/public/attachments/DOC-418460A1.pdf (data available at https://www.fcc.gov/sites/default/files/VTS_Historical_Data_Thru_D24.zip).

[4] Stephen J. Blumberg & Julian V. Luke, Nat’l Ctr. for Health Stats., Wireless Substitution: Early Release of Estimates from the National Health Interview Survey, July–December 2024 tbl. 1 (2025), https://www.cdc.gov/nchs/data/nhis/earlyrelease/wireless202506.pdf.

[5] AT&T Inc., Edited Transcript: 2024 Analyst & Investor Day (comments of Susan Johnson, EVP & GM, Wireline Transformation & Global Supply Chain, Dec. 3, 2024), https://investors.att.com/~/media/Files/A/ATT-IR-V2/reports-and-presentations/transcript-2024-12-03.pdf (hereinafter Johnson).

[6] See Opening Testimony of Dr. Mark Israel ¶¶ 19–33, A.23-03-003 (Cal. Pub. Utils. Comm’n Dec. 19, 2023).

[7] Pacific Bell Tel. Co. d/b/a AT&T Cal. (U 1001 C), Opening Comments on Administrative Law Judge’s Ruling Issuing Staff Proposal for Comment at 10, R.24-06-012 (Cal. Pub. Utils. Comm’n Jan. 30, 2026).

[8] AT&T Inc., Annual Report (Form 10-K) for the Fiscal Year Ended December 31, 2024 (Feb. 2025), https://investors.att.com/~/media/Files/A/ATT-IR-V2/financial-reports/t-2024-12-31-10k-final.pdf.

[9] AT&T Inc., Edited Transcript of Q4 2024 AT&T Inc. Earnings Call (Jan. 27, 2025), https://investors.att.com/~/media/Files/A/ATT-IR-V2/financial-reports/quarterly-earnings/2024/4Q24/t-usq-transcript-2025-01-27.pdf.

[10] See Jake Neenan, AT&T Approved to Discontinue Service at More Than 30% of Copper Footprint This Year, Broadband Breakfast (Jan. 13, 2026), https://broadbandbreakfast.com/at-t-approved-to-discontinue-service-at-more-than-30-of-copper-footprint-this-year.

[11] Verizon Commc’ns Inc., Edited Transcript: Investor Day 2022 (Mar. 3, 2022), https://www.verizon.com/about/sites/default/files/2022-03/Investor-Day-2022-Transcript.pdf.

[12] Sarah Thomas, Verizon Saves 60% Swapping Copper for Fiber, Light Reading (May 19, 2015), https://www.lightreading.com/cable-technology/verizon-saves-60-swapping-copper-for-fiber.

[13] Sean Buckley, Verizon Seeks FCC Permission to Retire Copper in 8 Markets, Emphasizes Call to Revise Processes, Fierce Network (Sept. 20, 2017), https://www.fierce-network.com/telecom/verizon-seeks-fcc-permission-to-retire-copper-8-markets-emphasizes-call-to-revise-processes.

[14] Id. See also Melissa Anders, John Vazquez Finds New Life in Old Buildings for Verizon, Am. Builders Q. (2015), https://americanbuildersquarterly.com/2015/verizon (“As modern telecommunications infrastructure migrates to fiber optics, away from copper wire, and from mechanical switches to digital ones, Verizon expects to reduce its technical footprint by as much as 80% where it makes these conversions, according to Vazquez.”).

[15] Lumen Techs., Inc., Financial Trending Schedule: 4th Quarter 2025 (2026), https://s21.q4cdn.com/756714007/files/doc_earnings/2025/q4/supplemental-info/Final-LUMN-Q4-2025-FTS.pdf.

[16] Press Release, AT&T, AT&T to Acquire Lumen’s Consumer Fiber-to-the-Home Business (May 21, 2025), https://about.att.com/story/2025/lumen-mass-markets-fiber-business.html.

[17] Press Release, Frontier Commc’ns, Fourth Quarter and Full Year 2024 Results (Feb. 20, 2025), https://investor.frontier.com/news/news-details/2025/Frontier-Reports-Fourth-Quarter-and-Full-Year-2024-Results/default.aspx.

[18] Press Release, Frontier Commc’ns, Third Quarter 2025 Results (Oct. 28, 2025), https://investor.frontier.com/news/news-details/2025/Frontier-Reports-Third-Quarter-2025-Results/default.aspx.

[19] See Press Release, Verizon, Verizon and Frontier: Regulatory Approval (Jan. 15, 2026), https://www.verizon.com/about/news/verizon-and-frontier-regulatory-approval.

[20] Cincinnati Bell Inc., Quarterly Report (Form 10-Q) (Nov. 12, 2025).

[21] Press Release, Brightspeed, Brightspeed Reaches More than 2M Fiber-Enabled Locations Across 20-State Footprint Ahead of Plan, Upsizes Build Goal to 5M+ (Apr. 2, 2025), https://www.brightspeed.com/brightspeed-news/Brightspeed_Reaches_More_than_2M_Fiber-Enabled_Locations.

[22] Linda Hardesty, Brightspeed Can Replace Copper with Unique Wireless Technology, Fierce Network (Sept. 9, 2024), https://www.fierce-network.com/broadband/brightspeed-can-replace-copper-unique-wireless-technology.

[23] See Jake Neenan, Consolidated Looking to Discontinue Copper at 45,000 New England Locations, Broadband Breakfast (July 10, 2025), https://broadbandbreakfast.com/consolidated-looking-to-discontinue-copper-at-45-000-new-england-locations; Jake Neenan, Consolidated Looking to Retire More Copper Phone Lines in New England, Broadband Breakfast (Aug. 19, 2025), https://broadbandbreakfast.com/consolidated-looking-to-retire-more-copper-phone-lines-in-new-england.

[24] Jake Neenan, FCC Approves Sale of Consolidated Communications to PE Firms, Broadband Breakfast (Dec. 9, 2024), https://broadbandbreakfast.com/fcc-approves-sale-of-consolidated-communications-to-pe-firms.

[25] Range reflects estimated energy savings of 166 kWh per subscriber (Altafiber) to 346 kWh per subscriber (AT&T).

[26] 340,000 MWh ÷ 983,000 decline in network access lines and DSL subscribers. See AT&T 10-K, supra note 8.

[27] AT&T, Energy Management (July 3, 2025), https://sustainability.att.com/priority-topics/energy-management.

[28] Rhonda Johnson, Building Networks for the Next Century, Not the Last One, AT&T Connects (May 15, 2024), https://www.attconnects.com/stories/building-networks-for-the-next-century-not-the-last-one.

[29] Nadja T., Are You Considering the Carbon Footprint of Your Internet Service?, Altafiber Blog (Oct. 14, 2024)., https://blog.altafiber.com/are-you-considering-the-carbon-footprint-of-your-internet-service.

[30] U.S. Bureau of Labor Statistics, Average Price: Electricity per Kilowatt-Hour in U.S. City Average (APU000072610), FRED, Fed. Rsrv. Bank of St. Louis, https://fred.stlouisfed.org/series/APU000072610 (last visited Feb. 24, 2026).

[31] Nadja T., Leaving a Greener Legacy, Altafiber Blog (Oct. 27, 2022), https://blog.altafiber.com/greener-legacy.

[32] Verizon, Verizon ESG Report 2023 (2024), https://www.verizon.com/about/sites/default/files/Verizon-2023-ESG-Report.pdf.

[33] Aislin Sullivan, Pushkar Tandon, Roshene McCool, Amalia Diaz & Constantin Herrmann, A Sustainable Future with Optical Fiber, Corning White Paper (Mar. 2023), https://www.corning.com/catalog/coc/documents/white-papers/WP1000.pdf.

[34] Telefónica, Connectivity Solutions’ Life Cycle Assessment: Executive Report (2022), https://www.telefonica.com/en/wp-content/uploads/sites/5/2022/03/connectivity-solutions-life-cycle-assessment.pdf.

[35] See Ramboll, Greener Connections: Understanding the Environmental Impacts of Fiber and Copper Communications Networks 3 (2025), https://ustelecom.org/wp-content/uploads/2025/01/Greener-Connections_Final.pdf (noting copper lines “are more susceptible to water damage than fiber optic lines” and fiber optic lines are “less susceptible to weather/climate-related events, such as flooding”).

[36] See Mike Robuck, AT&T Makes Case Against Keeping Copper, Mobile World Live (May 22, 2024). https://www.mobileworldlive.com/att/att-makes-case-against-keeping-copper.

[37] See Sullivan et al., supra note 33; Ramboll, supra note 35 (documenting fiber’s greater resilience and durability compared to copper).

[38] Johnson, supra note 5.

[39] Fiber Broadband Ass’n, Operational Expenses for All-Fiber Networks Are Far Lower Than for Other Access Networks (June 2020), https://fiberbroadband.org/wp-content/uploads/2023/03/Access-Network-OpEx-Analysis-White-Paper.pdf.

[40] Id.

[41] Ramboll, supra note 35.

[42] Pete Humes, How to Reduce Truck Rolls with Remote Visual Support, SightCall (Feb. 15, 2023), https://sightcall.com/blog/how-to-reduce-truck-rolls.

[43] Ramboll, supra note 35.

[44] Int’l Monetary Fund, Global Price of Copper [PCOPPUSDM], FRED, Fed. Rsrv. Bank of St. Louis, https://fred.stlouisfed.org/series/PCOPPUSDM (last visited Feb. 17, 2026).

[45] See Nicole Cobler, Copper Thefts in Dallas-Fort Worth Disrupt Utility and Emergency Services, Axios Dallas (Dec. 15, 2025), https://www.axios.com/local/dallas/2025/12/15/copper-thefts-dallas-fort-worth-att; see also Matt Egan, Copper Prices Are Rising. Thieves Are Taking Notice, CNN Business (Dec. 22, 2025), https://www.cnn.com/2025/12/22/economy/copper-wire-theft-att-outages.

[46] Susan Santana, Teaming Up to Tackle Copper Theft, AT&T Connects (July 16, 2025), https://www.attconnects.com/stories/teaming-up-to-tackle-copper-theft.

[47] USTelecom, Protecting the Nation’s Critical Communications Infrastructure from Theft & Vandalism (Oct. 2025), https://ustelecom.org/wp-content/uploads/2025/10/Protecting-Critical-Communications-Infrastructure-Report-Fall-2025.pdf.

[48] Id. The April 2025 edition of the report documented 5,770 incidents during the June–December 2024 period.

[49] Id.

[50] Egan, supra note 45; see also L.A. Bureau of Street Lighting, Outages and Issues, https://lalights.lacity.org/residents/outages_and_issues.html (last visited Feb. 18, 2026).

[51] FBI, Copper Thefts Threaten U.S. Critical Infrastructure (Dec. 3, 2008), https://archives.fbi.gov/archives/news/stories/2008/december/copper_120308.

[52] USTelecom, supra note 47.

[53] Egan, supra note 45.

[54] USTelecom, supra note 47.

[55] Id.

[56] Edward J. Lopez, The Real Costs of Communications Outages Due to Infrastructure Theft or Vandalism (Oct. 2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5526102 (also available at https://protectcci.org/wp-content/uploads/2025/10/Economic-Impact-Study_1001_2025.pdf).

[57] Western Electric/Lucent Modern Telephone Switching Systems, Telephone World (June 19, 2023), https://telephoneworld.org/telephone-switching-systems/western-electric-lucent-modern-telephone-switching-systems.

[58] Northern Telecom/Nortel Digital Central Office Switches, Telephone World (May 31, 2021), https://telephoneworld.org/telephone-switching-systems/northern-telecom-nortel-digital-central-office-switches.

[59] In re Nortel Networks Inc., 469 B.R. 478 (Bankr. D. Del. 2012); Long-Struggling Nortel Claims Bankruptcy, CBS News (Jan. 13, 2009), https://www.cbsnews.com/news/long-struggling-nortel-claims-bankruptcy.

[60] Application of Foothills Rural Tel. Coop. Corp. for Certificate of Convenience & Necessity, No. 2008-00163, Ex. B (Ky. Pub. Serv. Comm’n May 2, 2008), https://psc.ky.gov/PSCSCF/2008%20cases/2008-00163/Foothills_Application_050208.pdf (Siemens “Product Phase Out” notice listing the last possible order date and end-of-manufacturing date of July 15, 2007, for the EWSD core).

[61] Siemens EWSD Digital Electronic Switching System, Nokia Support Portal, https://customer.nokia.com/support/s/product2/siemens-ewsd-digital-electronic-switching-system/01t41000004gEmyAAE (last visited Feb. 17, 2026).

[62] See, e.g., Carritech, Lucent 5ESS, https://carritech.com/products/alcatel-lucent/lucent-5ess (last visited Feb. 17, 2026); Worldwide Supply, Nortel, https://worldwidesupply.net/brand/nortel (last visited Feb. 17, 2026); Carritech, Siemens EWSD, https://carritech.com/products/siemens-2/siemens-ewsd (last visited Feb. 17, 2026).

[63] Clayton Cummins, Tinker Retires Phone System Dating Back Over 40 Years, Tinker Air Force Base News (Oct. 24, 2024), https://www.tinker.af.mil/News/Article-Display/Article/3945606/tinker-retires-phone-system-dating-back-over-40-years.

[64] Robert Schult, Why Legacy Telecom Networks Are Giving Way to the Future, TeleGeography (Oct. 10, 2024), https://resources.telegeography.com/why-legacy-telecom-networks-are-giving-way-to-the-future.

[65] See, e.g., Cisco, End-of-Sale and End-of-Life Announcement for the Cisco ONS 15454 SONET/SDH SFP Short Haul Transceiver Module (Jan. 11, 2019), https://www.cisco.com/c/en/us/products/collateral/optical-networking/ons-15454-series-multiservice-provisioning-platforms/eos-eol-notice-c51-730654.html.

[66] Bo Gowan, SONET/SDH Is Dead–Really This Time, Ciena (June 19, 2013), https://www.ciena.com/insights/articles/SONETSDH-is-dead-really-this-time-prx.html.

[67] Johnson, supra note 5.

[68] Brattle Group, Fiber Deployment Has Significant Incremental Economic Benefits (2024), https://www.brattle.com/insights-events/publications/fiber-deployment-has-significant-incremental-economic-benefits-according-to-a-recent-brattle-report.

[69] Fiber Broadband Ass’n, supra note 39.

[70] WIK-Consult, Copper Switch-Off: European Experience and Practical Considerations (White Paper, Q3 2020), https://www.wik.org/fileadmin/Studien/2020/Copper_switch-off_whitepaper.pdf.

[71] Accenture, Economic Impact of Completing the Upgrade of nbn’s FTTN Network (2024), https://www.nbnco.com.au/content/dam/nbn/documents/about-nbn/reports/reports-and-publications/accenture-economic-impact-of-completing-fttn-upgrade.pdf.coredownload.pdf.

[72] See Blumberg & Luke, supra note 4 (reporting lower wireless-only adoption among adults 65 and older and rural-urban disparities in wireline deployment).

I. Introduction

Debates over the privacy implications of large language models (LLMs) often rest on an intuitive but unexamined premise: that these systems “store” personal data in the same way a database stores records. Regulators and commentators may therefore reason from database-based regulatory frameworks toward LLMs. The analogy has intuitive appeal but obscures both how language models function technically and how privacy risk arises in practice.

This issue brief argues that the database analogy misleads. Empirical research shows that memorization—verbatim or near-verbatim reproduction of training text—is rare relative to modern corpus scale, concentrated in low-entropy and highly duplicated material, and meaningfully mitigated through practices such as data curation, deduplication, and output filtering. LLMs do not maintain retrievable records. Their parameters encode probabilistic relationships across language, rather than tables of stored entries. Treating them as repositories of personally identifiable information risks imposing rules poorly aligned with both technical reality and observed risk.

Much confusion stems from definitions. In empirical machine-learning research, “memorization” refers to an output-observable event: reproduction of training text under specific prompting conditions. In privacy discourse, the term often serves as shorthand for broader concerns, including inference and hallucination. This issue brief uses “memorization” in the narrower empirical sense because it maps most directly onto what law can regulate—evidence of disclosure in outputs. Other risks, such as inaccurate statements about individuals, may raise separate legal questions, but they differ analytically from memorization and should not be treated as proof that a model stores personal data.

Framed this way, the central question is narrow: whether an LLM’s outputs can constitute a standalone privacy violation by disclosing identifiable personal information. That inquiry differs from the legality of training on datasets that include personal data and from a model’s ability to generate plausible statements through statistical inference. The analysis focuses on the model as a deployed system and on its outputs, because existing legal regimes generally attach liability to disclosure, misuse, or failure to safeguard information in defined contexts.

The discussion proceeds in four parts. Section II reviews the technical literature on memorization, explaining how researchers define, measure, and mitigate it and describing its frequency and distribution. Section III distinguishes research-relevant memorization from privacy-relevant harm and explains why hallucination and inference do not themselves demonstrate data leakage. Section IV evaluates how current law addresses these issues, surveying U.S. federal statutes and California’s California Privacy Rights Act to assess regulatory fit. Section V concludes.

Taken together, the evidence supports three propositions. Large language models are not databases, memorization of personal information is atypical, and privacy risk arises primarily at the point of output and use, rather than from internal statistical representations. Regulatory approaches that disregard these distinctions risk overdeterrence by constraining socially valuable technologies to address marginal or mischaracterized harms.

II. The Technical and Empirical Reality of LLM Memorization

Whether large language models (“LLMs”) memorize personal data in a way that constitutes a standalone privacy violation requires careful attention to both technology and law. Much of the debate turns on terminology. Engineers, empirical privacy researchers, and legal doctrine use “memorization” differently, which often obscures the relevant question.

For this issue brief, memorization refers to a specific technical event: a model reproduces training text verbatim or near-verbatim, typically under targeted or adversarial prompting. This concept is narrower than the legal category of personal-data processing.[1] LLMs ordinarily generate probabilistic reconstructions from learned token distributions, not stored records, and most outputs do not retrieve identifiable data.

The empirical literature further shows that memorization is conditional, rather than typical. It concentrates in repeated or low-entropy material, arises most often under contrived prompts, and appears at low rates relative to modern training-corpus size. Developers also deploy multiple safeguards—e.g., dataset curation, deduplication, red-teaming, filtering, and decoding controls—to reduce the likelihood of verbatim disclosure, though no technique eliminates the possibility across all prompts.

This section therefore reviews what the evidence demonstrates about memorization and how those findings inform a narrower legal inquiry: whether the models themselves, as distinct from the act of training or deployment, inherently violate privacy law.

A. What ‘Memorization’ Means in Large Language Models

In modern usage, “memorization” does not mean simply that a model has seen similar material before, nor that an output resembles its training data. Instead, memorization occurs when a model produces the same sequence of tokens—or a sufficiently similar sequence under standard string-distance metrics—that appears in the training data.[2] This definition reflects the dominant approach in empirical research on extraction and leakage. It is not the only one. Other work defines memorization through membership inference, generalization gaps, or differential privacy.[3] For present purposes, we focus on verbatim and near-verbatim reproduction because it most directly relates to output-stage privacy risk and to the legal question of whether a model functions as a repository of personal data.

Under this framework, large language models do not store information as records or tables. Their parameters encode probability distributions over token sequences, not discrete entries comparable to database rows.[4] Most outputs therefore reflect probabilistic reconstruction—novel text generated from learned distributions, rather than retrieved records.

Risk arises when the boundary between reconstruction and retrieval blurs. If a sequence appears frequently in the training data, or occurs in low-entropy boilerplate structures, the model may generate it with unusually high probability under carefully constructed extraction prompts, often continuation-style or otherwise adversarial. Empirical work demonstrates that such leakage is real and measurable.[5]

Defining memorization in practice remains difficult because natural-language documents commonly contain overlapping and repeated text. Material known to be in the training data can closely resemble material that is not, creating an inherently fuzzy boundary between members and non-members.[6] As a result, standard membership definitions may fail to capture relevant leakage. Membership-inference techniques may classify passages that are lexically or semantically similar to members as non-members, even when privacy auditors would still regard the disclosure as meaningful information leakage.[7]

B. Mitigating Memorization: Techniques and Tradeoffs

Frontier-model developers and the technical literature describe a range of measures designed to reduce verbatim memorization and the risk of disclosing sensitive material. The specific techniques—and the degree of public disclosure—vary across systems.[8]

Reported safeguards begin at training. Developers employ dataset curation and filtering,[9] deduplication to limit repetition-driven regurgitation,[10] and selective inclusion or exclusion of sources more likely to contain personal information. They also deploy post-training and operational controls, including red-teaming, classifier-based filtering, and decoding or sampling heuristics intended to reduce the likelihood that a model emits sensitive or low-entropy strings in response to ordinary prompts.[11] Researchers have also explored differential-privacy-enhanced training, although it generally involves meaningful tradeoffs in performance and utility.[12]

A separate line of research pursues more ambitious interventions, such as machine unlearning or targeted suppression of specific information in a trained model. These techniques remain technically immature.[13] Complex unlearning objectives can degrade model quality or fail to remove information embedded in distributed internal representations.

Mitigation therefore involves tradeoffs. Existing techniques can materially reduce observable memorization and the likelihood of verbatim disclosure, but they do not guarantee elimination of leakage across the full space of prompts, adversarial strategies, and distribution shifts, particularly at frontier scale.[14] More aggressive privacy-preserving approaches remain an active research area and can impose substantial costs in performance, training complexity, and practical utility.

C. Memorization Is Rare and Conditional

Across multiple studies, verbatim memorization appears uncommon relative to the scale of modern training corpora.[15] Importantly, these estimates do not measure “the fraction of the training set memorized.” They measure the share of test cases in which a model, given a short snippet, continues the passage by reproducing the original training text verbatim. Under that benchmark, even very large models trained on multi-trillion-token datasets show low reproduction rates—on the order of about 1–4% in recent frontier-model evaluations[16]—and newer model generations often copy less than earlier ones.[17]

These tests typically evaluate ordinary text. They do not directly answer a different policy question: whether a model can be induced to output sensitive personal information, which researchers assess using different targets and methods. Against trillion-token corpora, the total amount of text extractable verbatim, even with adversarial prompting, remains small. Scale cuts both ways. Larger models may increase copying risk, but improved curation and deduplication can offset it, and newer generations often combine greater scale with cleaner training data.[18]

The strongest extraction methods rely on contrived prompts such as “continue the following sequence exactly,” seeded with partial strings known to appear repeatedly in the corpus.[19] Without this scaffolding, random prompting produces little to no verbatim reproduction.[20]

An extraction study by Nicholas Carlini et al. illustrates the point.[21] In more than 600,000 model generations, the researchers confirmed only 604 verbatim extractions—about a 0.1% rate.[22] Even these required highly adversarial sampling and ranking procedures, as well as independent access to the same web text used to train GPT-2 to verify matches.[23] Extraction was therefore possible but rare: confirming memorization required prior access to the underlying data.[24]

A later large-scale analysis reached similar conclusions.[25] Across several major open-source models, only 0.03% to 1.4% of outputs contained recoverable training text, even after massive output generation and the deployment of specialized matching tools.[26] More sophisticated methods did not materially change the result. Extractable memorization exists, but it is a low-base-rate phenomenon.[27] Fine-tuned chat models require additional attack techniques and still produce memorized sequences at rates well below 1% of total output.[28]

Two empirical regularities emerge. First, duplication strongly predicts memorization.[29] Repetition is the most reliable indicator: strings appearing tens, hundreds, or thousands of times—such as license blocks, email signatures, news boilerplate, or SEO spam—are far more likely to be reproduced. Deduplicating training corpora can reduce verbatim memorization by an order of magnitude.

Second, low-entropy strings are disproportionately memorized. Highly predictable content—addresses, phone numbers, templated contact lines (“Contact me at …”), or schematic code—has few plausible continuations and is easier to reproduce verbatim.[30] High-entropy material, such as narrative prose or unique personal messages, rarely reappears in the same form.[31] Memorization risk is therefore unevenly distributed across data types.

Taken together, these patterns indicate that memorization is localized. It does not describe ordinary model behavior. It appears under specific conditions and with specific types of content.

III. From Memorization to Disclosure: The Legal Significance of Model Outputs

The literature contains an important definitional gap. Most machine-learning research uses “memorization” to mean verbatim reproduction of training text. Privacy-law debates often use the term more loosely, implying that a model internally retains personal information as a discrete object. Conflating these concepts obscures the relevant legal inquiry.

From a policy perspective, verbatim reproduction is neither necessary nor sufficient for privacy risk. A model may generate statements about an identifiable person—accurate or not—without reproducing training text verbatim. Conversely, a model may reproduce memorized strings that contain no personal information, such as boilerplate open-source license language.

For this issue brief, the analysis has two distinct components. First, whether personal data can meaningfully be said to reside in a model’s internal parameters. Second, and more legally significant, whether a model emits information in its outputs in a way that constitutes disclosure or misuse.

We focus on the second question. Accordingly, we set aside the permissibility of training on datasets that may contain personal data and instead examine whether a trained model, treated as a discrete object, can itself create a standalone privacy violation through its outputs. As the discussion of U.S. law below shows, liability generally attaches to disclosure, use, or access, not to the mere existence of statistical representations in model weights.

Under this framework, empirical memorization research—centered on verbatim reproduction—is an imperfect proxy for privacy harm. It identifies a measurable category of risk, but it does not capture every way personal information might appear in outputs. Privacy-relevant harm depends not only on memorized sequences, but also on whether information associated with identifiable individuals emerges under prompting, whether through direct regurgitation or transformations, such as paraphrase or translation.

A. Personal Data Outputs Are Not Necessarily Memorization

What does the evidence show about personal data specifically? The empirical record supports several propositions.

Researchers have demonstrated that personal phone numbers, email addresses, and short biographical details can sometimes be extracted from LLMs. These events, however, typically require that the information appears repeatedly online or in templated sources, e.g., faculty directories or scraped social-media biographies. The individuals most at risk therefore tend to have large digital footprints—public figures, academics, journalists, and others with SEO-dense online profiles. By contrast, models are unlikely to reproduce one-off personal information about private individuals, unless the information has been duplicated extensively across the web.

Many outputs that appear to contain personal data reflect probabilistic reconstruction, rather than verbatim retrieval from training data. In deployed systems, outputs that contain apparent personally identifiable information (PII) may also originate from user-provided prompts, uploaded documents, retrieval-augmented generation (RAG), or integrated web search, rather than training-data memorization. Often, the model generates a plausible statement based on statistical associations, rather than stored facts. When a model states that “John Doe is a lawyer in Chicago,” the output may be a hallucination or a statistical interpolation—common name plus common profession plus major city—or a paraphrased reconstruction. In ordinary usage, this does not resemble retrieval of a stored record.

The distinction matters. Treating probabilistic fabrication as equivalent to unauthorized disclosure of factual personal data would impose deterrence disproportionate to the underlying harm, particularly when the asserted personal data was never processed during training.

This leads to a broader point: hallucination is not evidence of memorization. Generative models routinely produce plausible but fabricated claims about individuals based on learned correlations. Equating those outputs with disclosure of training-set personal data collapses two distinct phenomena and treats speculative generation as proof of data leakage.

A regulatory approach that collapses hallucination into leakage would effectively treat statistical inference as the exfiltration of a stored record or the processing of personal data.[32] This concern differs from cases in which outputs combine publicly available information about real individuals with false or distorted claims, which instead resemble accuracy-based or defamation-style harms.[33] Conflating such hybrid outputs with memorization-driven leakage would distort incentives by encouraging developers to suppress useful generative capabilities to avoid liability for outputs that do not disclose stored records. The likely result would be reduced social utility, including diminished capacity for models to generate content, assist reasoning, or perform tasks that inherently involve interpolation.

Accordingly, the presence of PII in model outputs does not, by itself, show that a model memorized that information in a manner sufficient for a standalone privacy-violation claim. Under U.S. law, privacy liability generally turns on disclosure, unreasonable publicity, breach of confidentiality, or specific statutory triggers—not the existence of statistical associations within a model.

IV. How Existing Privacy Law Applies to Large Language Models

This section examines how different legal frameworks respond to output-disclosure risks associated with memorization, and how those frameworks map (or fail to map) onto LLM systems that often combine a base model with retrieval, user-supplied inputs, and other data sources.

The foregoing analysis has two immediate implications for policymakers. First, empirical evidence sharply limits the circumstances in which large language models plausibly resemble repositories of personal data. Memorization is rare, context-dependent, and most relevant at the point of output, rather than within internal parameters. Second, existing privacy regimes differ in how closely they track those technical realities.

This section examines how current legal frameworks respond to the memorization question and what that response reveals about regulatory fit. U.S. federal privacy law generally follows a sectoral, actor- and disclosure-based model: liability attaches when regulated entities collect, use, safeguard, or disclose identifiable information in defined relationships and contexts. California’s approach, while broader and more risk-oriented, similarly focuses on business practices, proportionality, and foreseeable harms, rather than the mere existence of statistical representations.

The goal is modest. This issue brief does not propose a comprehensive new privacy regime. Instead, it evaluates whether existing doctrines align with the technical characteristics of LLM systems and identifies where mismatches may produce over- or under-deterrence. The discussion first surveys federal privacy statutes and then turns to California’s hybrid framework, emphasizing how both regimes address output-related disclosure risks in systems that may combine a base model with retrieval tools, user-supplied inputs, and other external data sources.

A. Federal Privacy Statutes: Actor- and Disclosure-Based Liability

U.S. federal law generally follows a disclosure- and misuse-based conception of privacy, rather than a comprehensive data-protection model. Liability typically turns on whether a regulated actor disclosed, misused, or failed to safeguard identifiable personal information in defined contexts. None of the major federal statutes treat the transformation of text into statistical parameters as legally relevant storage, memorization, or processing. The legal trigger is misuse of identifiable data, not the existence of internal representations.

More broadly, U.S. privacy law—especially at the federal level—does not treat possession of statistically encoded information as per se unlawful. Privacy torts and related statutory provisions ordinarily require conduct such as public disclosure of private facts, intrusion upon seclusion, appropriation of a person’s name or likeness, or false light. Federal statutes including the Health Insurance Portability and Accountability Act (HIPAA), the Fair Credit Reporting Act (FCRA), the Children’s Online Privacy Protection Act (COPPA), the Gramm–Leach–Bliley Act (GLBA), and the Video Privacy Protection Act (VPPA) similarly define specific actors, specific data types, and specific prohibited acts. Across these regimes, liability attaches to identifiable information handled in regulated relationships and communicated to others—not to background exposure during model training.

This structure produces a consistent pattern across the statutes discussed below. Each is sectoral and context-dependent: HIPAA regulates health-care actors, FCRA governs consumer-reporting activities, COPPA addresses child-directed services, GLBA covers financial institutions, and the VPPA targets video-service providers. In each case, obligations arise at collection, use, disclosure, safeguarding, or eligibility decisionmaking. None treats a general-purpose model developer as regulated merely because a model trained on heterogeneous text, and none treats internal model parameters as legally cognizable records.[34]

Accordingly, the legally salient questions concern deployment and outputs. Statutory risk arises when a covered entity uses an LLM in a way that discloses identifiable information, incorporates protected data into eligibility determinations, or fails to implement required safeguards—not when training produces statistical representations that, standing alone, do not identify an individual. This issue brief surveys these frameworks to illustrate how U.S. privacy law evaluates conduct involving information systems and why those rules generally operate at the point of use and disclosure rather than at the level of model architecture.

1. HIPAA’s Actor-Based Limits

The Health Insurance Portability and Accountability Act (HIPAA) imposes privacy and security obligations on a defined set of regulated actors: “covered entities,” including health-care providers, health plans, and clearinghouses, as well as their business associates.[35] The statute protects identifiable “protected health information” (PHI),[36] and liability generally arises from the unauthorized use or disclosure of PHI within the health-care ecosystem.[37]

HIPAA does not regulate noncovered entities merely because they process information that contains health-related content. By extension, the statute ordinarily would not apply to statistical or machine-learning models that ingest text containing PHI when developed by a noncovered entity.[38] Absent covered-entity status or a business-associate relationship, a developer building a general-purpose LLM trained on heterogeneous web data is not subject to HIPAA and does not maintain PHI in a legally cognizable form. As a result, HIPAA generally does not reach an LLM’s internal numerical parameters, even if fragments of health-related text contributed to the training corpus.

Relevance to LLM Memorization: HIPAA is a sectoral, actor-based regime. It regulates who handles information and imposes duties tied to the use and disclosure of identifiable PHI within that regulated system. That structure makes HIPAA a poor fit for treating a general-purpose LLM’s internal parameters as a repository of PHI. The statute is more plausibly implicated at the output and deployment stage—e.g., when a covered entity or business associate uses an LLM in a way that results in an unauthorized disclosure of identifiable PHI—rather than at the training stage conducted by a noncovered developer.

2. FCRA’s Role- and Purpose-Based Limits

The Fair Credit Reporting Act (FCRA) governs a defined set of actors and activities within the consumer-credit ecosystem. It applies primarily to “consumer reporting agencies” (CRAs),[39] defined as entities that, for fees or on a cooperative nonprofit basis, regularly assemble or evaluate consumer information to furnish consumer reports to third parties.[40] A “consumer report” is any communication by a CRA bearing on a consumer’s creditworthiness, character, reputation, personal characteristics, or mode of living that is used, or expected to be used, to determine eligibility for credit, insurance, employment, or other statutorily specified purposes.[41]

Regulatory guidance reflects this role-based structure. The Federal Trade Commission (FTC) describes the statute as protecting information collected by credit bureaus, medical-information companies, and tenant-screening services, and emphasizes that such information may be furnished only for permissible purposes.[42] The Consumer Financial Protection Bureau similarly frames FCRA obligations around CRAs and users of consumer reports, including furnishers that supply information to CRAs and entities that rely on reports for eligibility decisions.[43] Academic commentary likewise characterizes FCRA as regulating the collection, maintenance, and disclosure of consumer information by CRAs and related participants in that reporting system.

Liability therefore arises when a CRA fails to meet the statute’s accuracy, disclosure, or permissible-purpose requirements, or when a furnisher or user violates duties tied to consumer reports—e.g., by supplying inaccurate data to a CRA or using a report without a permissible purpose. FCRA does not cover every entity that handles consumer information and does not regulate all databases containing personal data. It applies only when three elements align: a qualifying actor, a qualifying communication (a consumer report), and a qualifying use (an eligibility determination or another enumerated purpose).

Relevance to LLM Memorization: FCRA is a role- and purpose-based regime, not a general personal-data statute. A general-purpose LLM developer typically does not become a CRA merely because the model trained on heterogeneous text that includes consumer-related information. The statute instead becomes relevant at deployment. If an LLM system generates or supplies information that functions as a consumer report, or is used in an eligibility decision in a way that makes the operator resemble a CRA or covered user, FCRA’s accuracy, disclosure, and permissible-purpose requirements may apply. The key question is therefore not whether a model internally “contains” consumer data, but whether the system assembles, evaluates, and communicates consumer information for covered purposes.

3. COPPA’s Operator- and Context-Specific Scope

The Children’s Online Privacy Protection Act (COPPA) imposes obligations on a defined group of operators: those that run online services “directed to children” under age 13 or that have actual knowledge they collect personal information from a child.[44] The FTC explains that the law applies to operators of commercial websites and online services directed to children and also to general-audience services when the operator knows a child is providing personal information.[45]

COPPA regulates business practices surrounding the collection, use, and disclosure of children’s personal information. It requires parental consent, clear privacy notices, limits on data retention, and restrictions on sharing with third parties.[46] Its regulatory triggers attach to an operator’s handling of identifiable children’s data in providing the service. FTC guidance emphasizes that COPPA governs what information an operator collects, how it collects it, how it uses it, and how it discloses it—all tied to identifiable information associated with a child user.[47]

COPPA does not extend to downstream statistical or technical transformations once data no longer functions as identifiable personal information in the hands of a covered operator. A developer training a general-purpose LLM on broad web data, without operating a child-directed service or knowingly collecting children’s information from its own users, therefore falls outside the statute. In that setting, a model’s internal representations do not qualify as “personal information” under COPPA.

Relevance to LLM Memorization: COPPA is operator- and context-specific. Its duties attach to child-directed services and regulate the collection, use, retention, and disclosure of identifiable children’s data in providing that service. The statute fits front-end collection and deployment practices—consent flows, notices, retention limits, and onward disclosure—far better than it fits treating a general-purpose model’s parameters as children’s personal information. The most plausible COPPA issue therefore arises at the output stage: a covered operator could violate the statute by using an LLM to disclose a child’s identifiable information or by collecting such information through the interface without compliant consent and notice.

4. GLBA’s Financial-Institution Scope

The Gramm–Leach–Bliley Act (GLBA) governs the privacy and security practices of financial institutions, broadly defined as entities “significantly engaged” in offering financial products or services to consumers.[48] FTC guidance makes the boundary explicit: GLBA applies to businesses such as lenders, check-cashing services, and financial advisers.[49]

GLBA’s privacy framework centers on nonpublic personal information (NPI) arising from consumer financial relationships.[50] Financial institutions must provide privacy notices, limit disclosure of NPI to nonaffiliated third parties unless an exception applies, and implement administrative, technical, and physical safeguards to protect the information.[51] Implementing regulations focus on failures to safeguard NPI or improper disclosure of identifiable NPI, not on unrelated entities that merely process information touching on financial topics.[52]

Because GLBA obligations attach to financial institutions and to their handling of identifiable NPI, the statute does not extend to entities outside the financial-services context or to downstream statistical transformations of data occurring outside a regulated financial relationship. A general-purpose machine-learning model trained on web text does not become a financial institution simply because the corpus contains finance-related material. Likewise, statistical model parameters do not constitute “disclosure” or “sharing” of NPI under the statute.

Relevance to LLM Memorization: GLBA is a sectoral, entity-based regime focused on financial institutions’ treatment of customer financial information. Its core duties are operational: provide notices, limit certain disclosures of NPI, and maintain safeguards. The statute fits LLM deployment inside regulated institutions and their service-provider relationships, where the question is whether use of an LLM results in improper disclosure of identifiable NPI or inadequate safeguards. By contrast, GLBA does not plausibly treat a general-purpose model developer as a regulated financial institution or internal model parameters as NPI disclosure merely because training data included finance-related text. GLBA concerns arise when customer NPI enters, is processed by, or is output from an LLM in a manner that implicates safeguarding or disclosure obligations—not from background exposure during model training.

5. VPPA’s Disclosure-Based Scope

The Video Privacy Protection Act (VPPA) limits when a “video tape service provider” may disclose information identifying a consumer’s video-viewing history.[53] The statute defines such a provider as a business engaged in the rental, sale, or delivery of prerecorded video cassette tapes or similar audiovisual materials.[54] Courts have interpreted modern on-demand streaming services as falling within the statute’s scope.[55]

Liability arises when a covered provider knowingly discloses PII linking an individual to specific video materials or services.[56] As with other federal privacy statutes, the VPPA turns on disclosure of identifiable information by a covered provider, not on internal technical processing within a general-purpose model.

Relevance to LLM Memorization: The VPPA is a narrow, disclosure-focused statute aimed at a specific context: video-service providers and the release of information connecting an identifiable person to particular viewing choices. Its trigger is the knowing disclosure of protected identifiers tied to video materials or services, not abstract data use. The statute therefore becomes relevant to LLM systems only when deployed within, or on behalf of, a covered streaming provider in a way that could reveal an individual’s viewing history, e.g., through outputs, logs, or third-party sharing. By contrast, the VPPA does not plausibly treat a general-purpose model’s internal parameters as protected PII or treat training exposure to audiovisual-related text as disclosure of viewing history.

6. An Output- and Harm-Based Privacy Framework

The FTC’s “unfairness” authority is sometimes described as a more flexible federal privacy tool. Even so, liability under FTC doctrine requires a substantial consumer injury that consumers cannot reasonably avoid and that lacks countervailing benefits.[57] Internal model encodings do not meet that standard by themselves, because any injury depends on a downstream output. The mere existence of model weights capable of producing PII-like text therefore does not, standing alone, implicate the FTC’s privacy or data-security framework. Even the federal government’s broadest consumer-protection authority does not convert model parameters into regulated personal data.

Across the statutes surveyed, a consistent principle emerges. Liability arises when a system discloses identifiable information, is used in an eligibility decision, or otherwise enables unauthorized access or misuse in a regulated context.[58] The statutes attach duties to specific actors—health-care providers, consumer-reporting agencies, child-directed service operators, financial institutions, and video-service providers—and to defined activities such as collection, safeguarding, and disclosure. None treats a trained model’s internal statistical structure as a legally cognizable repository of personal data.

The same logic applies to memorization. The presence of memorized PII within model weights is insufficient, by itself, to create liability. Legal risk materializes when outputs reveal identifiable information under circumstances that make the disclosure unlawful or when deployment practices violate statutory duties. Even state regimes that extend somewhat further, such as the California Privacy Rights Act (CPRA, covered in Section IV.B), focus on business practices surrounding the handling of personal information, rather than the internal statistical architecture of a trained model.[59]

Taken together, U.S. privacy law reflects an output- and harm-focused framework. It evaluates how information is used, disclosed, or safeguarded in real-world interactions, not whether statistical parameters could theoretically encode personal data. Under that structure, treating LLM model weights as inherently regulated personal information is difficult to reconcile with existing doctrine.

B. The CPRA Regulates Data Practices and Outputs, Not Model Weights

California’s privacy regime, governed by the California Privacy Rights Act (CPRA)[60] and administered by the California Privacy Protection Agency (CPPA),[61] differs from federal law in structure but not in its basic trigger. The statute establishes baseline consumer rights, creates a category of “sensitive personal information,” and authorizes a dedicated regulator. It does not impose a general lawful-basis requirement for data processing. Instead, it regulates business practices surrounding collection, use, retention, sharing, and security. Liability turns on proportionality, consumer choice, and risk, not on the mere existence of internal representations derived from personal data. The CPRA also excludes certain publicly available and lawfully obtained truthful information that is a matter of public concern, which affects how the statute treats widely disseminated information about public figures.[62]

This structure matters for memorization. The CPRA does not expressly treat machine-learning model weights as personal information, nor does it presume statistical parameters are regulated simply because personal data contributed to their formation. The statute defines personal information relationally—data that identifies, relates to, describes, or can reasonably be linked to a particular consumer or household. Its exclusion for publicly available and newsworthy information further narrows the set of regulated outputs. Internal parameters therefore fall outside the statute, absent reasonable linkability or output-stage disclosure.

Recent California developments nonetheless move beyond notice-and-opt-out compliance and toward system-oriented governance relevant to memorization risk. The CPPA emphasizes data minimization, limiting collection, use, retention, and sharing to what is reasonably necessary and proportionate to disclosed purposes.[63] Although framed in legal rather than technical terms, minimization addresses upstream conditions associated with memorization, including excessive retention, overbroad ingestion, and large-scale duplication of low-entropy personal data. In practice, the requirement encourages data curation and deduplication that can reduce verbatim or near-verbatim reproduction.

California has also adopted governance requirements that create compliance touchpoints for advanced AI systems, without relying on a database analogy. CPPA rules on cybersecurity audits and risk assessments require certain businesses to evaluate and document foreseeable privacy and security risks and to implement proportional safeguards.[64] These obligations are model-adjacent, rather than model-centric. They do not assume that an AI model stores personal data in the ordinary sense. Instead, they focus on whether system design, deployment, and safeguards reasonably address identifiable risks, including the possibility that sensitive information could appear in outputs under foreseeable use or misuse.

V. Conclusion

This issue brief advances a narrower, empirically grounded understanding of memorization in large language models and urges greater care in mapping that concept onto privacy law. The technical literature consistently shows that memorization—verbatim or near-verbatim reproduction of training text—is rare relative to modern corpus scale, concentrated in low-entropy and highly duplicated material, and meaningfully reduced through established safeguards such as data curation, deduplication, and output filtering. Memorization is not a general feature of model behavior. It is a localized phenomenon that appears under specific conditions, often requiring targeted prompting.

U.S. privacy law largely tracks this reality. Across privacy torts and sector-specific statutes, liability attaches to disclosure, misuse, or failure to safeguard identifiable personal information in defined relationships and contexts. The law does not treat internal statistical representations as personal data merely because personal information may have contributed to their formation, and it does not equate probabilistic inference or hallucination with the exfiltration of stored records. A model’s weights, standing alone, do not establish a privacy violation. What matters is whether identifiable information is actually disclosed, accessed, or used in a legally relevant way at the output or deployment stage.

California’s framework is broader but points in a similar direction. The California Privacy Rights Act introduces data minimization, sensitive personal information, and risk-assessment obligations that expand oversight beyond federal sectoral statutes. Even so, California law remains focused on business practices, proportionality, and downstream sharing, rather than on treating models as repositories of personal data. In practice, the statute creates additional compliance obligations for AI systems without converting internal parameters into regulated records.

Several implications follow for policymakers. First, large language models are not databases. Their parameters encode probabilistic relationships rather than discrete entries, and memorization of personal information is atypical. Second, hallucination and inference raise different concerns from memorization and should not be treated as evidence of data leakage. Third, regulatory approaches that rely on database analogies or categorical assumptions about model behavior risk imposing disproportionate costs relative to the privacy harms at issue.

Accordingly, broad classification rules that treat all generated personal data—real or fabricated—as retrieved from storage risk overdeterrence. They encourage unnecessary restrictions on generative systems even when the risk of disclosing memorized personal information is low. A more coherent approach would align legal obligations with empirical realities by focusing on context, risk, and outputs. Such an approach better protects privacy, while preserving the substantial social value generative models can provide.

[1] See Regulation (EU) 2016/679 of the European Parliament and of the Council (General Data Protection Regulation) art. 4(2), 2016 O.J. (L 119) 1 (defining “processing” broadly to include “any operation or set of operations” performed on personal data, including collection, storage, use, disclosure, or deletion).

[2] See Gemma 3 Technical Report, arXiv:2503.19786, at 9 (Mar. 25, 2025), https://arxiv.org/abs/2503.19786 (defining “exact memorization” as token-for-token reproduction and “approximate memorization” as matching within a 10% edit distance); see also The Llama 3 Herd of Models, arXiv:2407.21783 (July 31, 2024), https://arxiv.org/abs/2407.21783 (discussing tests that probe whether models can reproduce training-data text verbatim).

[3] See Nicholas Carlini et al., The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks, in Proceedings of the 28th USENIX Security Symposium (2019), https://www.usenix.org/conference/usenixsecurity19/presentation/carlini; Vitaly Feldman & Chiyuan Zhang, What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation, arXiv:2008.03703 (Aug. 9, 2020), https://arxiv.org/abs/2008.03703; Wanrong Zhang et al., Leakage of Dataset Properties in Multi-Party Machine Learning, in Proceedings of the 30th USENIX Security Symposium (2021), https://www.usenix.org/conference/usenixsecurity21/presentation/zhang-wanrong.

[4] See Michael Duan et al., Do Membership Inference Attacks Work on Large Language Models?, arXiv:2402.07841 (Sept. 16, 2024), https://arxiv.org/abs/2402.07841 (explaining that the model is an autoregressive language model that predicts the probability distribution over the next token given a prompt).

[5] See Nicholas Carlini et al., Extracting Training Data from Large Language Models, in Proceedings of the 30th USENIX Security Symposium (2021), https://arxiv.org/pdf/2012.07805; Milad Nasr et al., Scalable Extraction of Training Data from (Production) Language Models, arXiv:2311.17035 (Nov. 28, 2023), https://arxiv.org/abs/2311.17035.

[6] Duan et al., supra note 4, at 6-7.

[7] Id. at 8-9.

[8] See Badrinath Ramakrishnan & Akshaya Balaji, Assessing and Mitigating Data Memorization Risks in Fine-Tuned Large Language Models, arXiv:2508.14062, at 3, 5, 41–42 (Aug. 10, 2025), https://arxiv.org/abs/2508.14062; Kunj Joshi et al., Randomized Masked Finetuning: An Efficient Way to Mitigate Memorization of PIIs in LLMs, arXiv:2512.03310, at 6–7 (Feb. 9, 2026), https://arxiv.org/abs/2512.03310.

[9] Gemma 3 Technical Report, supra note 2, at 2-3.

[10] The Llama 3 Herd of Models, supra note 2, at 4-5, 53-54.

[11] The Llama 3 Herd of Models, supra note 2, at 47, 49; Gemma 3 Technical Report, supra note 2, at 2-3.

[12] See Carlini, supra note 5, at 2644.

[13] See A. Feder Cooper et al., Machine Unlearning Doesn’t Do What You Think: Lessons for Generative AI Policy and Research, arXiv:2412.06966 (Dec. 9, 2024), https://arxiv.org/abs/2412.06966v2.

[14] See Carlini et al., supra note 5 (noting some techniques “help mitigate memorization but cannot prevent” it entirely); Da Yu, Differentially Private Fine-tuning of Language Models, arXiv:2110.06500v2 (July 18, 2022), https://arxiv.org/pdf/2110.06500.

[15] The Llama 3 Herd of Models, supra note 2, at 1.

[16] Id. at 41-42; Gemma 3 Technical Report, supra note 2, at 9.

[17] Id.

^[18] Alexander Xiong et al., The Landscape of Memorization in LLMs: Mechanisms, Measurement, and Mitigation, arXiv:2507.05578, at 2 (Dec. 12, 2025), https://arxiv.org/abs/2507.05578.

[19] See Gemma 3 Technical Report, supra note 2, at 9 (describing memorization tests that use a fixed prompt structure with a 50-token prefix and 50-token suffix to measure exact and approximate memorization).

[20] See The Llama 3 Herd of Models, supra note 2, (noting that effective memorization audits rely on targeted, structured prompts rather than random inputs); Jamie Hayes et al., Strong Membership Inference Attacks on Massive Datasets and (Moderately) Large Language Models, arXiv:2505.18773v1, at 2 (May 24, 2025), https://arxiv.org/html/2505.18773v1 (explaining that unstructured prompts rarely elicit verbatim training text and that extraction tests instead target high-frequency sequences likely to be overfit); id. (describing sampling prompts and expected outputs based on their frequency in the training corpus); see also Gemma 3 Technical Report, supra note 2, at 8 (describing contrived prompt structures—e.g., a fixed 50-token prefix used to predict a 50-token suffix—to measure extractable memorization).

[21] See Carlini, supra note 5.

[22] Id.

[23] Id. at 13.

[24] Id. at 7.

[25] See Nasr, supra note 5.

[26] Id. at 7, 19.

[27] Id.

[28] Id.

[29] Id. at 14-15; Duan, supra note 4, at 4-5.

[30] See id. at 3, 6 (describing high overlap and repetition in domains such as GitHub); The Llama 3 Herd of Models, supra note 2, at 4, 43 tbl. 24 (discussing filtering of PII and code data); Nils Lukas et al., Analyzing Leakage of Personally Identifiable Information in Language Models, arXiv:2302.00539, at 346 (Feb. 1, 2023), https://arxiv.org/abs/2302.00539 (examining privacy risks from potential PII leakage).

[31] Duan et al., supra note 4, at 8 (noting that repetition and common phrasing are inherent features of natural-language data, making truly unique sequences rare).

[32] Relatedly, if “hallucination” refers to a model’s failure to reproduce accurate information, restricting the use of personal data in training is not an obvious fix. In some settings, adding high-quality, lawfully obtained data (including information about real-world entities) may improve calibration and reduce certain hallucinations, while raising separate privacy issues involving lawful basis, purpose limitation, and data minimization. In short, hallucination risk and training-data privacy risk are distinct; conflating them can lead to data restrictions that do not address the underlying failure mode.

[33] See NOYB – European Center for Digital Rights, ChatGPT Provides False Information About People, and OpenAI Can’t Correct It (Apr. 29, 2024), https://noyb.eu/en/chatgpt-provides-false-information-about-people-and-openai-cant-correct-it.

[34] Because, as discussed infra, these regimes regulate defined actors and the use or disclosure of identifiable information in specified contexts. Treating model weights as regulated “personal information” would stretch statutory definitions and equate statistical parameters with maintained records. That move would likely draw challenges as inconsistent with statutory text, structure, and historically understood triggers, and as an agency assertion of major new regulatory authority absent clear congressional authorization. See Loper Bright Enters. v. Raimondo, 603 U.S. 369 (2024), https://www.supremecourt.gov/opinions/23pdf/22-451_7m58.pdf.

[35] See Covered Entities and Business Associates, U.S. Dep’t of Health & Hum. Servs., https://www.hhs.gov/hipaa/for-professionals/covered-entities/index.html (last visited Feb. 17, 2026).

[36] Id.

[37] 42 U.S.C. § 1320d-6; see also Scope of Criminal Enforcement Under 42 U.S.C. § 1320d-6, U.S. Dep’t of Justice, Off. Legal Counsel (2005), https://www.justice.gov/sites/default/files/olc/opinions/attachments/2014/11/17/hipaa_final.htm.

[38] See Nicolas Terry, Protecting Patient Privacy in the Age of Big Data, Ind. Univ. Robert H. McKinney Sch. of L. (Sept. 27, 2012), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2153269, at 21–22 (explaining that health information protected by HIPAA in the hands of a covered entity may become unprotected when obtained or used by a noncovered entity, such as a data-analytics firm).

[39] 15 U.S.C. § 1681(b).

[40] 15 U.S.C. § 1681a(f).

[41] 15 U.S.C. § 1681a(d).

[42] 15 U.S.C. §§ 1681–1681x.

[43] Fair Credit Reporting Act, Consumer Fin. Prot. Bureau, Consumer Laws & Regulations (2012), https://files.consumerfinance.gov/f/documents/102012_cfpb_fair-credit-reporting-act-fcra_procedures.pdf.

[44] 15 U.S.C. § 6502.

[45] See Complying with COPPA: Frequently Asked Questions, Fed. Trade Comm’n, https://www.ftc.gov/business-guidance/resources/complying-coppa-frequently-asked-questions (last visited Feb. 16, 2026).

[46] 15 U.S.C. §§ 6502–03 (setting requirements for parental consent, notice, data minimization, and limits on use and disclosure of children’s information).

[47] See Children’s Online Privacy Protection Rule: A Six-Step Compliance Plan for Your Business, Fed. Trade Comm’n, https://www.ftc.gov/business-guidance/resources/childrens-online-privacy-protection-rule-six-step-compliance-plan-your-business (last visited Feb. 16, 2026).

[48] 16 C.F.R. § 313.3(k)(1); 15 U.S.C. § 6809(3)(A).

[49] 16 C.F.R. § 313.3(k)(1).

[50] See How to Comply with the Privacy of Consumer Financial Information Rule of the Gramm-Leach-Bliley Act, Fed. Trade Comm’n, https://www.ftc.gov/business-guidance/resources/how-comply-privacy-consumer-financial-information-rule-gramm-leach-bliley-act (last visited Feb. 16, 2026).

[51] Id.

[52] 16 C.F.R. pt. 314.

[53] 18 U.S.C. § 2710.

[54] 18 U.S.C. § 2710(a)(4).

[55] See In re Hulu Priv. Litig., 86 F. Supp. 3d 1090, 1095 (N.D. Cal. 2015); Buechler v. Gannett Co., No. CV 22-1464-CFC, 2023 WL 6389447, at *2 (D. Del. Oct. 2, 2023).

[56] 18 U.S.C. § 2710(a)(3).

[57] 16 C.F.R. § 424.1.

[58] See 15 U.S.C. §§ 45(a), 45(n) (requiring substantial consumer injury for unfairness liability); 18 U.S.C. § 1030(a)(2) (imposing liability for unauthorized access to obtain information); 42 U.S.C. § 1320d-6(a); 45 C.F.R. § 164.502(a) (prohibiting unlawful use or disclosure of protected health information); 18 U.S.C. § 2701(a) (imposing liability for unauthorized access to stored communications). These statutes generally attach liability to unauthorized access, use, or disclosure of identifiable information, not to the mere existence of internal data representations.

[59] See How CPRA Defines Personal Information, Transcend (May 19, 2023), https://transcend.io/blog/cpra-personal-information; Cal. Civ. Code § 1798.140 (defining “personal information” as information that identifies, relates to, or could reasonably be linked to a consumer); id. §§ 1798.105, 1798.110 (requiring disclosure of categories of personal information collected and the purposes for which it is used, regulating business handling of personal information rather than internal technical representations not reasonably linkable to a consumer).

[60] Cal. Civ. Code § 1798.199.100 (2025).

[61] Cal. Civ. Code § 1798.199.10 (2025).

[62] Cal. Civ. Code § 1798.140(v)(1)(L).

[63] See Applying Data Minimization to Consumer Requests, Cal. Priv. Prot. Agency, Enf’t Div., https://cppa.ca.gov/pdf/enfadvisory202401.pdf (last visited Feb. 15, 2026).

[64] See William E. Ridgway et al., California Finalizes CCPA Regulations for Automated Decision-Making Technology, Risk Assessments and Cybersecurity Audits, Skadden (Oct. 3, 2025), https://www.skadden.com/insights/publications/2025/10/california-finalizes-cppa-regulations.

TL;DR

Background: America’s telecommunications landscape has moved from copper-based “plain old telephone service” (POTS) to fiber-optic and wireless networks. Yet the Federal Communications Commission (FCC) still enforces legacy rules—particularly under Sections 214 and 251 of the Communications Act—that require carriers to keep aging copper lines in service.

In recent research, Eric Fruits and Brian Albrecht of the International Center for Law & Economics (ICLE) show these requirements force providers to maintain networks for a small, shrinking subscriber base. As customers leave copper, per-user costs rise, distorting investment and competition. 

But… The policy ignores consumers’ revealed preferences. Americans have largely abandoned landlines for mobile and Voice over Internet Protocol (VoIP) services. By 2024, about 79% of adults lived in wireless-only households, while fewer than 1% relied exclusively on landlines. Maintaining a parallel, obsolete network acts as a hidden tax on infrastructure investment, diverting capital and skilled labor from expanding fiber and other modern technologies that could help close the digital divide.

Moreover… Keeping copper networks creates a public-safety risk. As the global price of copper climbed toward $6 per pound in early 2026, theft of telecommunications wiring surged. Each incident costs an average of $8,735 to repair and can trigger outages that disrupt 911 systems, hospital communications, and military facilities. By compelling carriers to maintain copper lines, regulators leave a ready supply of valuable metal in the ground—effectively subsidizing organized theft and undermining the public-safety goals the rules aim to advance.

KEY TAKEAWAYS

Consumers Have Moved On

The FCC’s most recent “Voice Telephone Services” report shows how quickly the market has changed. From 2014 to 2024, subscribers using switched-access lines fell 77%, from nearly 73 million to just over 16 million. Copper local-loop—or “last-mile”—connections declined even faster, dropping 81% from almost 66 million to about 12.5 million.

Copper’s share of the remaining wireline market also shrank. In 2014, 60% of wireline subscribers relied on copper last-mile connections; by 2024, only 28% did. Over the same period, mobile subscriptions rose from 322 million to 391 million, a 21% increase.

The Centers for Disease Control and Prevention’s (CDC) National Health Interview Survey confirms the consumer response. In 2024, about 79% of adults lived in wireless-only households, while just 0.9% lived in landline-only households.

Copper’s Hidden Costs

Maintaining a legacy network is not just a cost of doing business—it materially drains revenue. AT&T reported spending about $6 billion in 2023—roughly 5% of total revenue—to keep its copper network operating. In California alone, the company spends about $1 billion per year to serve customers who now account for roughly 3% of households in its service territory.

Much of the expense reflects copper’s fragility. Unlike fiber, copper is vulnerable to moisture and corrosion, requiring frequent “truck rolls” (maintenance dispatches) that cost $150 to $500 per visit. Verizon, by contrast, found that migrating customers to fiber cut maintenance dispatches 60%, producing about $180 million in annual operational savings. In 2025, the International Center for Law & Economics (ICLE) filed comments with the FCC noting that these savings could fund next-generation deployment rather than continued “firefighting” of deteriorating 20th-century technology.

Copper’s Energy Penalty

Legacy copper systems use far more energy than modern networks. Transmitting electrical signals over long distances and powering aging central-office equipment requires substantial electricity. Altafiber (formerly Cincinnati Bell) reported that its copper network uses about 172 kWh per subscriber each year, compared with 6 kWh for fiber—a 97% reduction in energy intensity.

AT&T estimates its shift from copper to fiber saved 340,000 megawatt-hours of electricity in 2024 alone. A Ramboll report similarly finds fiber at least 100 times more energy-efficient than copper during operation. Requiring carriers to keep copper networks running conflicts with environmental and net-zero goals, as legacy switches consume eight to 10 times the energy of the servers that replace them.

Obsolete Hardware, Rising Risk

Delayed retirement also means relying on obsolete hardware. Major switching systems, including the Lucent 5ESS and the Nortel DMS-100, have not been manufactured for decades. Nortel filed for bankruptcy in 2009, and Siemens issued “Product Phase Out” notices for its digital switches as early as 2007.

Carriers now scour secondary markets for refurbished parts. In one case, Tinker Air Force Base had to buy replacement switching components on eBay because they were no longer commercially available. Dependence on this gray market creates reliability risks and raises labor costs, as fewer technicians still possess the specialized knowledge needed to maintain legacy equipment.

Let Copper Retire

ICLE’s 202 comments to the FCC urged the agency to use its Section 10 forbearance authority. ICLE asked the commission to waive network-change notice rules under Section 251(c)(5) and service-discontinuance requirements under Section 214 where competitive alternatives exist. The comments argued current rules create unnecessary transaction costs and deadweight loss—economic value destroyed by regulation rather than market failure.

International evidence points the same way. A 2020 WIK-Consult study for the FTTH Council Europe found notice periods of up to five years significantly delayed copper retirement even where fiber was widely available, and recommended shorter timelines where coverage exists. An Accenture analysis for Australia’s government-owned broadband operator projected AU$10.4 billion in cumulative GDP gains from 2026 to 2034 from completing its fiber upgrade.

Fruits and Albrecht note in their ICLE issue brief that remaining copper subscribers—often older and more rural—deserve attention during any transition. The FCC’s proposed framework addresses this through notice and the availability of wireless and Voice over Internet Protocol (VoIP) alternatives. The policy question, the brief explains, is not whether transition costs exist but whether they justify spending billions each year to maintain infrastructure serving a shrinking customer base that is already leaving voluntarily.

For more on this issue, see Eric Fruits and Brian Albrecht’s ICLE issue brief “Paying to Stand Still: Legacy Copper Mandates in a Fiber World” and ICLE’s August 2025 comments to the FCC on “Reducing Barriers to Network Improvements and Service Changes.”

A bevy of states are racing to mandate “digital choice” in social media. The new bills promise easy data portability and forced interoperability among platforms—letting users carry their accounts, contacts, and content across services through open protocols. Utah enacted the first such law in 2025, and legislatures in Virginia, South Dakota, New York, California, and New Hampshire are now considering similar measures in their 2026 sessions.

The pitch sounds simple: give users control over their information. A closer look tells a different story. The bills never identify a clear market failure. Their interoperability mandates expose nonconsenting users to significant privacy risks. Their artificial-intelligence (AI) provisions do not cohere into a workable regulatory scheme. And lawmakers are moving ahead despite little evidence that consumers actually want social-media interoperability.

Read the full piece here.

The Federal Trade Commission (FTC) just announced a “landmark” settlement with one of the nation’s largest pharmacy benefit managers (PBMs). The problem is that it doesn’t actually end the PBM case—and it raises as many questions as it answers.

In its settlement with Express Scripts Inc. (ESI) and its affiliated entities, the agency says the deal will force fundamental changes to ESI’s business practices, increase transparency, and reduce patients’ out-of-pocket drug costs—including insulin—by as much as $7 billion over 10 years. Perhaps it will, although they don’t show their work on that one. It also promises new revenue for community pharmacies and aligns, according to the FTC, with the Trump administration’s health-care priorities.

But here’s wrinkle number one: the FTC settled only part of a larger case.

On Sept. 20, 2024, the FTC filed a complaint against the three largest prescription-drug benefit managers—Caremark Rx, Express Scripts (ESI), and OptumRx—along with their affiliated group purchasing organizations (GPOs). The agency alleged anticompetitive and unfair rebating practices “that have artificially inflated the list price of insulin drugs, impaired patients’ access to lower list price products, and shifted the cost of high insulin list prices to vulnerable patients.”

Here’s the redacted complaint.

The new agreement resolves the claims against ESI and its affiliates alone. The cases against the other PBMs remain pending, and the FTC has said nothing yet about how—or whether—those will be resolved.

Read the full piece here.

Regulators keep searching for a simple test to declare broadband markets “competitive.” The California Public Utilities Commission’s Public Advocates Office (Cal Advocates) thinks it found one: count gigabit networks. If a market lacks multiple overlapping gigabit-capable systems, Cal Advocates suggests in a report released last month, regulators should treat it as effectively noncompetitive.

That framing misses how competition actually operates. The report centers on premium gigabit speeds, even though most households neither need nor choose to buy them. Consumers typically prefer cheaper, sub-gigabit tiers that easily satisfy standard broadband definitions and common uses like streaming, videoconferencing, and gaming.

The pricing analysis compounds the problem. The study compares average prices for the lower-speed tiers in competitive markets to the highest-priced gigabit tiers in markets with only one gigabit provider, while assuming consumers in the latter subscribe to premium plans. In practice, the plans most households actually purchase—lower-speed tiers—show relatively stable pricing, regardless of gigabit overlap.

Even so, the report illustrates a broader regulatory trend: advocates increasingly rely on artificially high speed thresholds to label markets uncompetitive. That label often supports utility-style regulation and opposition to mergers, despite the basic economics of network industries. The enormous fixed costs of broadband infrastructure naturally limit how many terrestrial networks a single market can sustain.

At the same time, competition does not depend solely on duplicate fiber lines. Fiber, fixed wireless, and satellite technologies now discipline pricing and expand consumer choice. Prices have trended downward as these alternatives spread.

Policies built on flawed competitive benchmarks carry risks. If regulators suppress prices or block efficiency-enhancing consolidation based on mismeasured competition, providers lose the expected returns needed to finance upgrades and rural deployment. The result would slow investment, weaken network expansion, and harm the very consumers the policies aim to protect.

Read the full piece here.

Debates about federal preemption in artificial-intelligence (AI) policy often pose a stark choice: Congress adopts a national framework and states lose the ability to police harmful conduct, or states retain broad authority and businesses face a 50-state compliance patchwork that chills innovation. Our review of state AI bills suggests the debate is aimed at the wrong target. Most state legislative activity does not regulate how AI systems are built.

To clarify terms: an AI “model” is the core software—a set of mathematical parameters (“weights”) trained on large datasets—that produces text, images, predictions, or other outputs. An AI “system” is the broader product that uses one or more models, along with interfaces, data pipelines, guardrails, and other application logic. State bills use both terms, often loosely.

Regulating the “model layer” means imposing obligations on developers about how the core software is designed, trained, or evaluated. Regulating the “use layer” means governing what people and organizations do with AI systems once they exist—how they are deployed and what conduct they enable. Roughly three-quarters of the state AI bills we examined focus on downstream uses, rather than model design or development. As a result, they are unlikely to be preempted.

A federal framework could therefore focus on model-layer obligations while preserving traditional state police powers over uses such as fraud, deception, impersonation, election manipulation, and discrimination. Under that structure, most state laws would likely survive in some form.

Two important caveats apply.

First, labeling a bill as regulating “uses” does not necessarily make it a traditional police-powers measure. A statute targeting fraud that imposes joint-and-several liability on model developers for downstream misuse may read like a conduct rule. In practice, it functions as a model-layer obligation because it forces developers to change how systems are built, not merely how they are used. States considering use-layer legislation should ask whether their proposals effectively require design changes at the foundational level. Requirements that operate this way raise the same interstate-commerce and compliance-fragmentation concerns as direct model-layer regulation.

Second, even well-intentioned use-layer rules can burden interstate commerce. A state’s label does not control the constitutional analysis. Courts look to substance, not characterization.

Our survey does not predict how courts would resolve any particular case. No descriptive coding can. It does clarify the practical stakes: fears that federal AI preemption would eliminate most state AI law do not match what states are actually enacting and proposing.

Read the full piece here.

Antitrust has always been a strange regulatory enterprise. Businesses are largely free to engage in various commercial practices involving price, output, product design, distribution, research, and innovation—until they’re not. Outside the paradigmatic examples of explicit agreements among competitors to fix price and output, many business practices live in a gray zone. Whether a particular pricing practice, rebate structure, or distribution strategy is lawful cannot be known with certainty ex ante.

Instead, legality often turns on an ex post judicial inquiry under the “rule of reason,” where a court weighs competitive harms against procompetitive justifications and decides whether the conduct undermines the competitive process and harms consumers. Recurring pressures also push antitrust toward broader social goals, including worker employment levels and environmental impact.

Consider Epic Games’ lawsuit against Apple over various iOS policies, including the requirement that iPhone apps be distributed exclusively via the App Store. Apple adopted this distribution model when the App Store was launched in 2008, just one year after the iPhone entered the market in 2007. Over the next decade, the iPhone grew in popularity to become the leading smartphone in the United States.

In 2020, 12 years after the policy’s adoption, Epic Games sued and alleged the policy violated the Sherman Act due to “Apple’s substantial market power.” Although Apple ultimately prevailed in 2023 after a full trial and appellate review, it took three years of litigation to settle the question. The episode underscores an antitrust oddity: business practices implemented at entry can be retroactively recharacterized as unlawful once market success changes a firm’s market position.

Read the full piece here.

The fight over how to protect children online under the First Amendment has intensified in the opening weeks of the year. Two product-liability trials recently began asking whether social-media services such as Instagram and YouTube qualify as unsafe products for minors because of allegedly defective design features. At the same time, federal and state lawmakers have introduced bills that would impose liability on app stores if they fail to verify users’ ages and obtain parental consent before minors download apps.

Debate has largely centered on who should bear responsibility for harms to children online: the app stores or the apps themselves. The more important question, though, is when liability makes sense at all.

This post examines that question through a law & economics lens, with particular attention to H.R. 6333, the Parents Over Platforms Act (POPA). Sponsored by Reps. Jake Auchincloss (D-Mass.) and Erin Houchin (R-Ind.) the bill cleared the House Energy and Commerce Subcommittee on Commerce, Manufacturing, and Trade in December and now awaits consideration by the full committee.

Read the full piece here.

Policymakers and commentators often treat large language models (LLMs) as if they were searchable repositories of personal data. The intuition is understandable: these systems train on massive corpora that may include personal information, and they occasionally generate outputs referencing real people.

But the analogy is still wrong. And policy built on it risks distorting both innovation and privacy enforcement.

I’ve written a new issue brief examining the empirical evidence on LLM memorization, distinguishing it from analytically separate phenomena such as hallucination and inference, and surveying how existing U.S. privacy law addresses these issues. The research points in a consistent direction: large language models do not store personal data like databases, memorization of personal information is atypical, and privacy risk arises primarily at the point of output and use, not from internal statistical representations.

Read the full piece here.

Gail Slater has left the building. And that’s a shame, I think.

On Feb. 12, Slater posted the following on X:

It is with great sadness and abiding hope that I leave my role as AAG for Antitrust today. It was indeed the honor of a lifetime to serve in this role. Huge thanks to all who supported me this past year, most especially the men and women of @justiceatr

That was a short tenure. The Senate had confirmed her only 11 months earlier. Some reports said she resigned or “stepped down” (here and here). Others suggested she had been fired (here and here).

If you want inside details, you’ll have to look elsewhere. I don’t have them. But tension between the Antitrust Division and U.S. Justice Department (DOJ) leadership has been public since at least late July 2025, when reports said two senior division officials—Roger Alford and Bill Rinner—had “been fired for insubordination,” and not by Slater (see here and here).

Read the full piece here.

Europäische Wettbewerbsbehörden haben einen bewährten Weg gefunden, um aufmerksamkeit zu erregen: die Konfrontation mit amerikanischen Technologieunternehmen. Jed neue untersuchung bringt Schlagzeilen, politische anerkennung und institutionelles ansehen. Die Kosten dafür trägt jedoch nicht das Silicon Valley, sondern europäische Haushalte.

Read the full piece here.

Federal video regulation still treats broadcast, cable, and streaming as separate worlds. Consumers do not. The gap between how the law classifies video services and how people actually watch them is widening, and it increasingly distorts competition across the modern media marketplace.

Last week, the International Center for Law and Economics (ICLE) hosted a panel on the future of video competition in the United States. The conversation ranged widely—from broadcast-ownership caps and retransmission consent to smart-TV intermediaries, data-driven advertising, and the role of user-generated content. A consistent theme emerged: the legal silos structuring federal video regulation no longer reflect consumer experience. As Congress and the Federal Communications Commission (FCC) revisit video-competition policy in the coming year, those tensions will become harder to ignore.

For casual observers, the debate may appear limited to recent FCC broadcast-rule enforcement or high-profile streaming mergers. More attentive followers may point to broadcast-ownership reform or consolidation among major streaming firms. These developments, though, are only surface manifestations. As the video marketplace evolves, regulators confront a growing set of distinct but interrelated policy questions.

At the center of these changes is a simple fact: consumers no longer care about technical distinctions among cable, broadcast, and streaming. They just want to watch the content they choose. The FCC’s legal categories, however, were built for a 20th-century broadcast environment. They do not map cleanly onto how video is marketed, discovered, and consumed today. Updating regulation requires understanding how consumers perceive substitutable services and how siloed rules across technologies can impede competition on the merits.

Many of these distortions do not stem from recent FCC action, but from the Communications Act’s statutory architecture. Title III’s broadcast-licensing regime and Title VI’s cable-specific mandates were designed for spectrum scarcity and infrastructure bottlenecks that no longer define the marketplace. Piecemeal administrative action cannot resolve these structural tensions. A durable solution will require congressional engagement, not merely agency reinterpretation.

Read the full piece here.

The 1970s were a strange time, to put it mildly.

Chicken farmers gassed, drowned, and suffocated roughly a million baby chicks. “It’s cheaper to drown ’em than to put ’em down and raise ’em,” one Texas farmer explained. Dairy farmers slaughtered cows. Hog farmers culled breeding stock.

Why did any of this happen? Good old price controls.

This isn’t another “price controls are bad” post. (They are.) But a new paper of mine with Alex Tabarrok and Mark Whitmeyer adds something genuinely new: a theorem explaining why price controls generate exactly this kind of “chaos,” as we call it, and a new way to measure the costs without assuming a demand curve.

Most people, of course, think first about gasoline lines associated with 1970s price controls. I’ve written before about the 1970s gas crisis—odd-even rationing, fights at filling stations, and lines that vanished almost overnight when controls ended. In Maryland and Connecticut, lines stretched for miles. More than 90% of stations in Connecticut and Massachusetts rationed fuel. Some ran out entirely.

What gets less attention is the other side. In Idaho, Montana, Utah, and Wyoming, not a single surveyed station reported any problem, according to AAA survey data presented to President Ford during the crisis. Zero. Texas, the Deep South, and the Great Plains were, as Time magazine put it, “virtually awash with gasoline.”

So why would a 9% national gasoline shortfall produce more than 90% of stations rationing in Connecticut and no shortages in Idaho?

The answer points to the same mechanism that leads farmers to destroy livestock, rather than gradually scale back output. The key to understanding it, as always, is price theory.

Read the full piece here.

The neo-Brandeisian movement—emphasizing market structure and fairness over consumer welfare—has struggled to gain traction in several jurisdictions. In Europe, by contrast, lawmakers have codified its core premises. The Digital Markets Act (DMA) operationalizes this approach by prioritizing ex ante structural interventions intended to reshape rivalry in digital markets.

The regime’s most significant feature lies not in its high-level goals, but in its enforcement tools. Through “specification proceedings” under Article 8(2), the European Commission has begun to move from policing competitive constraints to directing system design. The Commission’s recent actions involving Google and the Android ecosystem illustrate that transition most clearly.

Read the full piece here.

Canada is on the verge of hard-coding costly mistakes into its merger policy.

The Competition Bureau’s proposed merger guidelines aim to translate Parliament’s recent overhaul of the Competition Act into enforcement practice. The draft instead risks entrenching a merger regime that treats market structure as destiny, discounts consumer-benefiting efficiencies, and substitutes blunt presumptions for evidence of competitive harm. The stakes extend beyond Canada, as jurisdictions worldwide reconsider how aggressively to police mergers.

The guidelines implement statutory changes adopted in 2023 and 2024. Those amendments eliminated Canada’s efficiencies defence, introduced structural presumptions tied to concentration and market share, authorized consideration of labor-market effects, and expanded enforcement authority across the board.

In comments submitted to the bureau this week, scholars at the International Center for Law & Economics (ICLE) identified three features of the draft that should give competition authorities pause.

Read the full piece here.

Germany’s competition watchdog has turned a familiar retail feature into an antitrust offense. In a recent decision, the German Federal Cartel Office (FCO) effectively faulted Amazon for showing shoppers only “price-competitive” offers on its German site.

At first glance, the case looks narrow—a dispute over the mechanics of the “Buy Box,” reminiscent of European actions against so-called “most-favored-nation” (MFN) clauses. In substance, it signals something broader. The decision illustrates two recurring problems in contemporary European competition policy.

First, the FCO privileges speculative theories of harm over observable market effects. Second, national authorities increasingly pursue aggressive ex ante interventions that risk fragmenting the EU’s single market, rather than harmonizing it.

Read the full piece here.

America’s financial system has always thrived when innovation, competition, and consumer empowerment work hand in hand. And when America’s financial system thrives, so too do Americans—from business owners to factory workers and families of all stripes. The long-anticipated, and soon to be released Consumer Financial Protection Bureau’s (CFPB’s) “Personal Financial Data Rights” rule—also known as the Section 1033 of Dodd-Frank—is an opportunity for the Trump administration to unleash a new wave of innovation that will empower consumers and make life more affordable for all Americans.

Read the full piece here.

Nearly 17 months after the Federal Trade Commission filed suit against the nation’s three largest pharmacy benefit managers (PBMs), the agency has reached a settlement with one of them: Express Scripts. The complaint alleged that PBMs harmed competition and patients by inflating insulin prices.

PBMs negotiate drug prices with pharmaceutical manufacturers and pharmacies on behalf of private health plans and public programs, such as Medicare and Medicaid. They leverage pooled purchasing power across dozens of plans to secure discounts and determine which drugs appear on plan formularies, as well as their order of preference. CVS Caremark, Express Scripts, and Optum Rx—the PBMs named in the FTC’s complaint—are vertically integrated with major insurers: Aetna, Cigna, and UnitedHealth Group, respectively. Together, they account for roughly 80% of U.S. prescription volume.

Health plans that lack vertical integration with a PBM can still access this negotiating leverage by contracting with one of the large PBMs. Although some PBMs charge flat administrative fees, compensation typically depends on manufacturer rebates tied to formulary placement. That structure can “steer” patients toward higher-rebated drugs over lower-priced competitors.

Read the full piece here.

Last month, the U.S. House Judiciary Committee released an interim staff report examining CVS Health’s relationships with digital pharmacy services. Led by Chairman Jim Jordan (R-Ohio), the report takes aim at a core policy concern: rising health-care costs. It frames competition and innovation as central tools for addressing those costs, a welcome emphasis in a debate often dominated by structural suspicion rather than market effects.

At the same time, the report leaves key questions unresolved. It gestures toward the responsibility of health-care firms to combat fraud and signals ambivalence about the role and value of pharmacy benefit managers (PBMs) in the marketplace, without fully engaging the tradeoffs those intermediaries present.

The timing is notable. Scrutiny of PBMs is intensifying, and the Federal Trade Commission (FTC) just announced a significant settlement with Express Scripts Inc. over insulin pricing, underscoring the broader regulatory pressure shaping the report’s backdrop.

Read the full piece here.

Brazil’s competition authority (CADE) and Apple signed a Dec. 29 settlement agreement (Termo de Compromisso de Cessação, or TCC) resolving a high-stakes antitrust investigation Mercado Livre initiated in 2022 in Brazil and Mexico. Mercado Livre is Latin America’s leading e-commerce and marketplace platform.

The agreement marks a watershed moment for Brazil’s digital economy. CADE will require Apple to restructure its iOS ecosystem by enabling third-party app stores, allowing alternative payment processors, and removing anti-steering restrictions. Through traditional ex post antitrust enforcement, CADE secured changes that legislative proposals—such as Bill 4.675/2025—would impose through ex ante regulation.

This post analyzes the settlement’s legal framework, CADE’s controversial market definition, the remedies’ proposed fee structure, and the agreement’s broader implications for the global digital-competition debate.

Read the full piece here.

Over the last few decades, antitrust scholars and practitioners have scrutinized the role of platforms—particularly intermediaries—in the internet economy. Many intermediary platforms also compete in the markets they facilitate. That dual role raises familiar concerns about self-preferencing and the risk that a firm may advantage its own products or services.

Regulators have focused primarily on large technology firms that develop mobile operating systems, app stores, and online marketplaces. The European Union’s Digital Markets Act (DMA) imposes obligations on statutorily defined “gatekeepers.” In the United States, Congress has floated parallel proposals, including the American Innovation and Choice Online Act (AICOA) and the Open App Markets Act (OAMA). Commentators have dissected these measures at length, often faulting their thin treatment of consumer effects and their break from traditional economic analysis of allegedly anticompetitive conduct.

Video markets have likewise drawn scrutiny for decades. Broadcast-ownership caps, retransmission-consent rules, and streaming mergers remain live policy flashpoints as consumers migrate to streaming as a primary viewing mode. Fights over market definition and consumer welfare persist, illustrated by the proposed Warner Bros. Discovery–Netflix transaction and consolidation involving broadcasters Tegna and Nexstar.

The television itself has escaped comparable attention.

For most of its history, the television was straightforward hardware—a display for external video inputs. Sets bundled broadcast receivers and ports for cable boxes, gaming consoles, DVD and Blu-ray players, and other peripheral devices.

That paradigm has shifted. Smart TVs—internet-connected sets that host downloadable media apps—now dominate the market. Aside from periodic litigation over user-privacy practices, regulators have largely left Smart TV platforms outside the platform-governance frame. That omission may prove temporary as the sector expands.

Smart TV platforms deliver many of the same procompetitive benefits seen in other digital intermediaries. Vertical integration across the television technology stack enables original-equipment manufacturers (OEMs) to cut retail prices and improve the precision and quality of television advertising.

Read the full piece here.

When competition authorities expand their legal toolkits, the most consequential policy choices often do not appear in the statute. They emerge later—in enforcement guidelines, presumptions, and priorities that determine how aggressively agencies will deploy new powers. Canada now finds itself squarely in that phase.

Recent amendments to the Competition Act illustrate the shift. Reforms enacted in 2023 and 2024 added “excessive and unfair selling prices” as a form of abuse of dominance, extended civil review to agreements between non-competitors, expanded private access to the Competition Tribunal, and sharply increased penalties. Corporations now face fines of up to C$25 million, rising to C$35 million for repeat violations, or three times the benefit gained.

These statutory changes have pushed interpretive authority downstream. The Competition Bureau must now give operational meaning to open-ended concepts and decide how forcefully to pursue them. Comments submitted by several scholars, including a recent set from the International Center for Law & Economics (ICLE), underscore how much turns on those implementation choices, rather than on statutory text alone.

Canada’s approach combines three elements that many jurisdictions continue to debate: expanded substantive theories of harm; broader private rights of action paired with monetary remedies; and dramatically higher penalties. Together, they create a real-world test of what happens when lawmakers broaden enforcement authority without strengthening the doctrinal guardrails meant to limit error costs and over-deterrence.

That experiment deserves close attention in Washington, Brussels, and other capitals. The underlying risks—legal uncertainty, inflated error costs, and pressure to turn competition law into de facto price regulation—are not uniquely Canadian. They reflect structural tensions in modern competition policy, especially when enforcement discretion expands faster than doctrinal discipline.

Read the full piece here.

Ahead of tomorrow’s Senate Judiciary Antitrust Subcommittee hearing, a group of former federal antitrust enforcers sent an open letter to the U.S. Justice Department (DOJ) and the full Judiciary Committee urging a consumer-welfare-focused review of the proposed Netflix–Warner Bros. Discovery merger. The letter rejects the progressive analytical framework advanced during the prior administration and calls for a return to established antitrust principles.

Read the full piece here.

Brightline rules promise clarity. The early enforcement record suggests something closer to friction.

Part I of this two-part series examined how mobile app-store anti-steering policies—rules that restrict developers from directing users to alternative offers or payment portals outside the app store—affect competition, consumers, and innovation. It also compared those policies with the restrictions upheld in Ohio v. American Express Co., in which the U.S. Supreme Court approved American Express’ anti-steering rules in a multisided credit-card transactions market.

Part II turns to enforcement. It compares how Apple’s App Store anti-steering policies and transaction fees have been treated in the Europe Union under the Digital Markets Act (DMA) and in the United States under federal and state antitrust law. The analysis focuses on what those divergent approaches have delivered so far for competition, commercial and legal certainty, enforcement costs, and the rule of law. Taken together, the early results raise serious questions about whether U.S. policymakers should continue pressing to import DMA-style ex ante regulation into American competition policy.

Read the full piece here.

Two years ago, the European Union’s flagship Digital Markets Act (DMA) took effect. The DMA promised to make platform markets “fairer and more contestable” for the businesses that rely on them to reach consumers by imposing a set of brightline mandates on designated gatekeepers.

In the United States, advocates of DMA-style competition rulemaking include former Federal Trade Commission (FTC) Chair Lina Khan and Sen. Amy Klobuchar (D-Minn.), whose American Innovation and Choice Online Act (AICOA)—which Congress has not enacted—closely tracks the DMA’s approach. Supporters argue that brightline rules lower enforcement costs and provide greater certainty for regulators and firms than case-by-case antitrust litigation against large technology companies. U.S. antitrust law, by contrast, generally requires enforcers and private plaintiffs to prove that a defendant possesses and has abused market power and that the challenged conduct produces net anticompetitive effects. Brightline regimes such as the DMA aim to bypass these showings altogether.

The ongoing battles between mobile app developers and competition authorities over Apple’s so-called anti-steering policies—fought in U.S. courts and before European regulators—put that case for digital platform rulemaking under strain. These disputes suggest that rigid rules do not necessarily simplify enforcement or align with the interests of the very developers who have lobbied for them.

This article examines app-store policies that restrict developers’ ability to inform users about, or direct them to, alternative payment options outside an app store’s proprietary in-app payment system, such as transactions completed on a developer’s own website. Part II compares how those policies fare under the DMA’s brightline, ex ante prohibitions with their treatment under the more flexible, effects-based framework of U.S. federal and state antitrust law.

Read the full piece here.

Japan rarely rushes to follow global regulatory fashion. But when it comes to mobile platforms, it has done exactly that. With the Mobile Software Competition Act (MSCA) taking effect in December 2025, Tokyo has chosen to regulate smartphones the European way — by laying down detailed rules in advance, rather than enforcing competition law case by case. The move may feel modern and decisive, but it may also be a costly mistake.

The MSCA is modeled on the European Union’s Digital Markets Act, though it is narrower in scope and will be enforced by the Japan Fair Trade Commission (JFTC), an agency known for consultation and restraint. That softer tone matters. But tone cannot fix a law built on shaky assumptions.

Read the full piece here.

Abstract

Price controls kill the incentive for arbitrage. We prove a ChaosTheorem: underabinding price ceiling, suppliers are indifferent across destinations, so arbitrarily small cost differences can determine the entire allocation. The economy tips to corner outcomes in which some markets are fully served while others are starved; small parameter changes flip the identity of the corners, generating discontinuous welfare jumps. These corner allocations create a distinct source of cross-market misallocation, separate from the aggregate quantity loss (the Harberger triangle) and from within-market misallocation emphasized in prior work. They also create an identification problem: welfare depends on demand far from the observed equilibrium. We derive sharp bounds on misallocation that require no parametric assumptions. In an efficient allocation, shadow prices are equalized across markets; combined with the adding-up constraint, this collapses the infinite-dimensional welfare problemtoaone-dimensionalsearchoveracommonshadowprice,withextremal losses achieved by piecewise-linear demand schedules. Calibrating the bounds to stationlevel AAA survey data from the 1973–74 U.S. gasoline crisis, misallocation losses range from roughly 1 to 9 times the Harberger triangle.

Abstract

Rapid technological change poses a continuing challenge for competition policy. Just as the emergence of digital platforms and ecosystems prompted departures from traditional antitrust enforcement and a recent wave of regulatory intervention, the rise of AI is again reshaping the competitive landscape, raising questions about the adequacy of existing analytical approaches and regulatory tools. Examining the expansion and role of AI-enabled assistants and agents, this paper analyzes strategies for integrating AI applications. It argues that AI’s disruptive potential requires a context-sensitive approach, rather than a mechanical replication of the Big Tech–centric framework that has characterized recent competition policy in digital markets.

I. Introduction

Artificial intelligence (AI) is rapidly becoming a general-purpose technology with immediate implications for competition policy.[1] As deployment spreads across search, commerce, productivity software, and communications services, AI tools are beginning to reshape how firms compete, organize production, and interact with users. At the downstream level, foundation models enable new product categories and add functionality to existing products across digital markets.

Much of this change now centers on AI assistants and agents. These systems increasingly mediate users’ online activity and reconfigure the traditional web-browsing experience around AI-centered interaction.[2]

AI assistants—e.g., OpenAI ChatGPT, Google Gemini, Microsoft Copilot, Anthropic Claude, Meta AI Assistant, and Mistral Le Chat—are largely reactive, language-based systems. They interpret natural-language prompts and generate responses using probabilistic reasoning and, in some cases, tool-augmented capabilities. Their operation remains prompt-driven: users initiate the task and define its boundaries.

Alongside these AI assistants, early agentic systems—e.g., OpenAI Operator, Anthropic Claude’s Computer Use, Perplexity’s Buy with Pro, and Google DeepMind’s Project Mariner—operate differently. These systems initiate actions, interact with external software environments, and adjust behavior to pursue specified objectives. They can execute multi-step workflows and sustain continuous processes that extend beyond simple response generation. The difference is functional rather than semantic: assistants answer questions, while agents perform tasks.

The expansion of generative AI has rapidly increased both the number and diversity of such systems.[3] AI assistants and agents increasingly function as access points to digital services rather than standalone applications.[4] Through a single interface, users can search, shop, book services, retrieve documents, or interact with third-party tools without leaving the conversational environment. AI-enabled browsers, including OpenAI’s ChatGPT Atlas and Perplexity’s Comet, reinforce this shift.

Firms are integrating these capabilities through two primary strategies. Some partner with third-party AI developers; others build proprietary systems and embed them into existing ecosystems. This strategic divergence tracks the competitive relationship between incumbent platforms and AI entrants. New entrants seek distribution through widespread embedding across platforms, while incumbents deploy vertically integrated AI to retain users within their ecosystems.

Competition authorities have begun to scrutinize these developments. Integration may improve quality and lower transaction costs, but regulators fear it may also reinforce market power. Agencies have raised concerns about tying, foreclosure, and self-preferencing, particularly where AI tools are embedded into widely used services.[5] Jurisdictions that adopted ex ante digital-market regulations are already reassessing whether those frameworks adequately address AI-enabled conduct.[6]

Early disputes illustrate the emerging terrain. Amazon has sued Perplexity, alleging its agent accessed user accounts and masked automated activity as human browsing.[7] Elon Musk has threatened litigation accusing Apple of manipulating App Store rankings to favor OpenAI’s ChatGPT after Apple integrated the service into iOS devices.[8] Meanwhile, the Italian competition authority opened proceedings regarding Meta’s preinstallation of Meta AI within WhatsApp.[9]

This paper analyzes the competitive implications of integrating AI applications into digital ecosystems under both antitrust law and sector-specific regulation.

From an antitrust perspective, the core issues are not new. They resemble longstanding debates over vertical integration and self-preferencing in digital markets.[10] Existing competition law remains flexible enough to address exclusionary conduct. The real novelty lies elsewhere: AI markets currently exhibit rapid entry, experimentation, and technological uncertainty. The speed with which new services and firms have emerged complicates predictions about durable market power. Replicating the familiar anti–Big Tech framework risks misdiagnosing competitive dynamics in a still-fluid environment.

The same caution applies to digital-platform regulation. A Big Tech-centric regulatory approach produces asymmetric coverage. The European Union’s (EU) Digital Markets Act (DMA), for example, captures AI functionality integrated into designated core platform services,[11] yet standalone AI applications may fall outside its scope regardless of growth or competitive significance. The rise of agentic systems exposes the difficulty regulators face in anticipating technological trajectories.[12]

Policy therefore confronts two opposing risks: delayed intervention that permits anticompetitive conduct to entrench, and premature intervention that distorts competition in evolving markets.

The paper proceeds as follows. Section II examines business strategies for deploying assistive AI services and uses the Meta AI investigation to illustrate potential antitrust concerns arising from integration into core ecosystems. Section III analyzes the challenges AI technologies pose for recently enacted digital-market regulations. Section IV evaluates the limits of a Big Tech-centric framework and questions the assumption that AI markets will follow the historical path of earlier digital platforms. Section V concludes.

II. Competition and Distribution in AI Assistant Markets

AI assistants and agents have become a central competitive frontier in online markets. A growing share of users now rely on these tools to interact with digital systems, retrieve and synthesize information, automate tasks, and delegate bounded autonomous actions across platforms and services. The rapid cadence of model releases and the continuing race for performance leadership reflect both the importance of these applications and the intensity of rivalry in this sector.

Competition between incumbent technology firms and AI entrants increasingly turns on how these systems are adopted and distributed. Because AI assistants may displace or mediate traditional general-purpose search and other gateway services,[13] new entrants pursue rapid diffusion through embedding across multiple platforms. Incumbent firms instead confront a familiar make-or-buy decision: partner with external AI developers or build proprietary systems.

Each option presents tradeoffs. Partnerships allow incumbents to reach competitive parity quickly and reduce disruption risk by aligning with potential rivals. Over time, however, reliance on third-party technology can create strategic dependence on external providers and their commercial success.[14] Internal development preserves control but requires significant investment and technological uncertainty.

Observed market behavior reflects these incentives. Most incumbents have integrated proprietary AI functionality into existing ecosystems. Google has incorporated generative-AI features into Gmail, Google Docs, and Search. Meta has embedded AI assistants into WhatsApp, Messenger, and Instagram. Microsoft integrated Copilot into the Office suite, although its initial capabilities relied heavily on its partnership with OpenAI. Microsoft has since diversified Copilot’s architecture by combining internal and external models to reduce reliance on OpenAI.

Many partnerships instead focus on upstream AI inputs—cloud infrastructure, computing chips, training data, and technical expertise—or take the form of financial investment rather than distribution agreements.[15]

New entrants follow a different strategy. OpenAI’s launch of “apps in ChatGPT” allows users to run third-party applications inside conversations, supported by an open developer toolkit (the Apps SDK).[16] Early partners include Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow, with additional participants such as OpenTable, Peloton, Target, Tripadvisor, and Uber expected to join.

OpenAI has also pursued commerce integration. Companies including Etsy, Shopify, and Walmart allow users to browse and purchase products directly through ChatGPT.[17] A partnership with PayPal enables instant checkout through PayPal’s digital wallet, while PayPal processes merchant payments.[18]

Incumbents have responded in kind. Google has partnered with firms including Shopify, Etsy, Wayfair, Target, and Walmart to develop an open standard for agentic commerce. The system enables users to complete purchases through Search or the Gemini application without switching between apps or webpages.[19]

Competition authorities have expressed concern about both types of strategies.

First, regulators scrutinize partnerships between large platforms and AI developers.[20] These arrangements may generate pro-competitive benefits by giving entrants access to capital, distribution, and essential inputs such as specialized computing capacity. Authorities nonetheless worry that partnerships may neutralize emerging competitors—for example, through contractual restrictions that limit downstream competition—and thereby reinforce incumbent positions across the AI value chain. [21]

Second, agencies have focused on the integration of AI into core platform services.[22] Ecosystem integration can improve product quality and reduce transaction costs, but authorities warn it may create foreclosure risks. Regulators point to tying and self-preferencing theories of harm: platforms may condition access to core services on use of proprietary AI or advantage their own tools through preinstallation, exclusive integration, or interoperability limits. Such practices could restrict user choice and raise barriers to rival applications.

Taken together, skepticism toward both partnerships and internal development reflects a broader concern that AI may strengthen existing digital ecosystems rather than disrupt them.[23] Current policy debates therefore largely adopt a Big Tech–centric perspective.

The ongoing Meta AI investigation provides a useful case study for evaluating these issues and the competitive implications of foundation-model assistants embedded within established digital ecosystems.

A. Meta AI and WhatsApp: A Tying Case in AI Markets

In July 2025, the Italian competition authority (ICA) opened an antitrust investigation into Meta’s decision to preinstall Meta AI within WhatsApp, combining its messaging service with its proprietary AI assistant.[24] The authority emphasized that Meta AI appeared in a prominent interface position and was integrated into the WhatsApp search bar, allowing users to interact with the assistant without opening a separate chat.[25]

The ICA also noted limited user control over the feature’s visibility. Users could access competing AI services by initiating separate chats, but they could not remove the Meta AI interface elements.[26] The authority further identified uncertainty regarding training data. User interactions with Meta AI appeared to contribute to model training, except for private messages and instances in which users explicitly opted out in specific chats.[27]

The investigation therefore centers on alleged anticompetitive tying. In the ICA’s view, preinstallation and preferential placement may give Meta a competitive advantage in the AI-chatbot market by leveraging its position in consumer messaging services.[28] The authority’s concern is that Meta can steer WhatsApp’s large user base into the emerging AI market not through competition on the merits, but through product integration.[29]

The ICA also emphasized the potential interaction between distribution and data. If Meta trains its model using interactions generated by a dominant messaging service, user-base leverage and data accumulation may reinforce each other. This feedback loop could produce lock-in effects and reduce reliance on competing assistants.[30]

Despite the technological novelty of AI, the theory of harm is conventional. The case reflects a familiar vertical-integration framework rather than a new antitrust doctrine. Competition law has repeatedly addressed similar conduct in digital markets,[31] and existing rules do not require structural revision simply because the product now incorporates AI.

European precedent illustrates the point. In Microsoft, the General Court held that the ubiquity of a dominant operating system could foreclose competition in the tied software market.[32] Bundling software with a preinstalled operating system allowed the tied product “to benefit from the ubiquity of that operating system … which cannot be counterbalanced by other methods of distributing media players.”[33]

A decade later, in Google Android, the European Commission found that Google preserved and strengthened its dominance in general search by requiring device manufacturers to preinstall Google Search and Chrome as a condition for licensing the Play Store and by imposing contractual restrictions that locked Android into a Google-controlled ecosystem.[34] The Commission concluded that preinstallation created a status quo bias that reduced both manufacturers’ incentives to preinstall rival applications and users’ incentives to download them.

More recently, in Facebook Marketplace, the Commission determined that tying Facebook Marketplace to Facebook abused Meta’s dominant position because integration provided a distribution advantage that rival platforms could not replicate.[35] Embedding Marketplace in the Facebook interface ensured universal visibility,[36] and although users could adjust certain settings, doing so required multiple complex steps that limited practical effectiveness.[37]

U.S. law reflects similar concerns. The U.S. District Court for the District of Columbia held that Google secured default-search status through anticompetitive distribution agreements with browser developers, device manufacturers, and carriers.[38] The U.S. District Court for the Northern District of California likewise found that Google unlawfully tied access to the Play Store to the use of Google Play Billing.[39] The court barred Google from requiring Play Billing for distributed applications and from imposing contractual restrictions that conditioned payment, distribution, or access to the Play Store on exclusive or preferential treatment.

Against this background, the ICA can rely on a well-developed tying framework. The authority must establish that: (i) the tying and tied products are distinct; (ii) Meta holds a dominant position in the tying market; (iii) users lack a genuine choice to obtain the tying product independently of the tied product; and (iv) the conduct is capable of producing exclusionary effects.

That does not mean the outcome will mirror prior cases. As discussed below, the distinctive feature of AI markets is competitive uncertainty. Even if Meta attempts to leverage messaging dominance into AI services, success cannot be presumed. The key question is whether preinstallation grants an advantage that rivals cannot offset through alternative distribution, product quality, or innovation.

The investigation has also expanded. The ICA examined Meta’s October 2025 business terms, which prohibit providers from using the WhatsApp Business Solution when AI assistants constitute the primary service offered.[40] AI tools remain permitted for ancillary functions such as automated customer support. The ICA adopted interim measures,[41] and both the European Commission and the Brazilian competition authority opened parallel inquiries.[42]

Authorities worry the policy may restrict output, market access, or technical development by preventing rival assistants from reaching users through WhatsApp. Meta responds that general-purpose chatbots fall outside WhatsApp’s intended function as a communication tool between businesses and users and that supporting such systems would require substantial operational resources.[43]

Although this paper focuses on the integration of Meta AI into WhatsApp, the parallel investigations offer additional insight into market conditions. The ICA’s interim measures suggest a market with multiple entrants attempting to reach users.[44] Some are smaller providers without proprietary distribution channels—e.g., Ira, Luzia, Poke, Puch AI, and Zapia—while others, including ChatGPT, Copilot, and Perplexity, already possess alternative distribution pathways. The dispute therefore turns not only on dominance, but also on whether messaging platforms function as essential gateways to user access.

III. AI and Digital-Market Regulation

If AI complicates antitrust analysis, it poses an even sharper challenge for recent digital-market regulation. The same technological change that motivates regulatory intervention may also undermine it. Rapid advances in AI risk rendering newly enacted regulatory frameworks incomplete or outdated.

The DMA illustrates the problem. As part of the Act’s first review, the European Commission is consulting stakeholders on whether the framework adequately addresses AI-enabled services and whether the list of core platform services and related obligations requires revision.[45] The DMA—like other digital-market regimes—was not designed with AI in mind. Legislators instead targeted large online platforms designated as gatekeepers, defined as firms that serve as gateways for business users to reach consumers and can leverage advantages, particularly access to data, across markets.

AI services may fall within DMA-type regulation through two main channels. First, an AI provider could itself offer a core platform service and qualify as a gatekeeper. Second, AI functionality embedded within an already designated platform service becomes subject to the obligations governing that service. The result is a differentiated regulatory landscape that treats incumbent platforms and AI entrants differently.

For incumbents, integration brings AI features inside the DMA’s scope. The primary question becomes whether existing obligations can adapt to new technologies. The Commission’s ongoing specification proceeding concerning Google illustrates this dynamic. The proceeding seeks to ensure that Google provides third-party AI services access to Android operating-system features comparable to those available to its own services, consistent with the vertical-interoperability requirement of Article 6(7) of the DMA.[46]

Standalone AI providers face the opposite situation. The DMA may not apply at all, regardless of market significance, because standalone AI applications do not clearly fit within any enumerated core platform service. Unless regulators reinterpret existing categories—e.g., search engines, browsers, or virtual assistants—AI developers cannot be designated as gatekeepers solely based on AI offerings.[47] This structural gap is particularly salient for agentic systems, which may reshape market intermediation while remaining formally outside the regulation’s scope.[48]

Regulatory expansion, however, carries its own risks. AI markets remain unsettled in technology, business models, and market structure. Extending ex ante obligations to AI services may constrain experimentation in product design, functionality, and platform architecture. Policymakers would need to predict both the direction and speed of technological change in an environment characterized by rapid innovation and uncertainty.

Premature regulation therefore risks misjudging market power. Authorities may underestimate future competition or overestimate the durability of concentration. Either error could distort innovation incentives and consumer outcomes. The regulatory problem is thus symmetrical: waiting too long may allow harmful conduct to emerge, but acting too early may suppress competitive experimentation.

IV. Rethinking the Big Tech-Centric Framework

Both the Meta AI investigation and the ongoing debate over revising the Digital Markets Act (DMA) reflect a similar analytical pattern. Assessments of AI competition remain anchored in the same Big Tech–centric framework that has shaped digital-market policy for two decades.

Competition authorities recognize that AI has stimulated innovation and entry. At the same time, they worry that markets for foundation models may follow the early trajectory of platform markets.[49] In this view, economies of scale and scope, network effects, data feedback loops, and limited multi-homing could generate “winner-takes-most” outcomes and eventual market tipping.[50] The concern is therefore prospective: large technology firms might shape AI markets in ways that reduce future competition by leveraging existing advantages across layers of the AI stack.

The ICA’s Meta AI investigation illustrates this reasoning. The authority emphasizes that AI development requires substantial computing power, high-quality data, specialized labor, and investment capital.[51] Because large platforms control many of these inputs and operate vertically integrated ecosystems, regulators fear they can both resist disruption and extend market power into adjacent markets.[52]

Sound policy analysis, however, must compare AI systems with prior digital platforms, rather than assume they are equivalent.[53] The literature suggests important differences.[54]

Network effects appear weaker for foundation models than for traditional platforms because individual users do not gain meaningful value from additional users. Data feedback loops exist, but the strategic importance of proprietary data may be declining as datasets become more abundant and synthetic data more common. Meanwhile, investment patterns show persistent entry: new AI firms have attracted substantial funding across successive rounds,[55] indicating both investor confidence and expectations of future competition.

In practice, many predicted entry barriers have proven less significant than anticipated. The diversity of downstream uses also makes universal tipping unlikely. Instead of a single dominant platform, AI services increasingly specialize across applications. Rapid market expansion and continued entry therefore challenge assumptions underlying current policy debates and suggest ongoing competitive pressure.

Some scholars further note that static measures of market power—such as market shares and margins—may understate competition in innovative industries.[56] Dynamic indicators, including investment levels, innovation rates, and the ability of smaller firms to attract capital, show meaningful opportunities for entry and expansion.[57] The emergence of firms such as OpenAI and Anthropic is difficult to reconcile with a theory that incumbents can seamlessly extend dominance into AI markets.[58]

Empirical evidence also complicates the narrative that data advantages determine competitive success. To date, large platforms have not translated existing data holdings into decisive superiority over AI startups. OpenAI’s ChatGPT, for example, has become the most widely used chatbot. By February 2025, it exceeded 400 million monthly active users[59] and accounted for roughly 86% of global chatbot traffic between April 2024 and March 2025.[60] By September 2025, it ranked among the world’s most visited websites, [61] and by November 2025 it was the most downloaded generative-AI mobile application.[62] Developer adoption likewise remains high.[63]

These figures nonetheless remain fluid. As of January 2026, ChatGPT’s traffic share had declined to roughly 65% amid growing competition from Gemini.[64] Anthropic’s Claude has also expanded rapidly and is projected to reach break-even earlier than OpenAI.[65] The relevant point is not which firm leads, but how quickly leadership can change.

Rapid turnover weakens predictions of durable concentration. Even unsuccessful entry can discipline incumbents in contestable markets.[66] The threat of displacement may therefore matter as much as actual market shares.

Against this background, a presumption against integration strategies by large technology firms risks analytical error. The concern resembles longstanding skepticism toward vertical integration. Yet vertical integration can generate efficiencies, eliminate double marginalization, reduce transaction costs, and improve product quality and coordination.

Applying traditional anti–Big Tech reasoning to AI disregards the current competitive environment. The market exhibits entry, diverse business models, rapid innovation, and abundant capital. Large platforms do not appear to hold a decisive advantage.

The Meta AI case illustrates the point. It is uncertain that integrating Meta AI into WhatsApp would materially harm competition in AI assistants, especially given the success of rivals such as ChatGPT. ChatGPT achieved rapid adoption through cross-platform integration and partnerships enabling users to shop, book services, and perform other tasks within a single interface. By contrast, Meta AI’s market share remained minimal—about 0.2% during April 2024–March 2025[67] and below 1% in January 2026[68]—and developer adoption was low.[69]

Industry practice further undermines a categorical rule against integration. Firms routinely use their own services as distribution channels for AI tools. In recent U.S. litigation, Judge Amit Mehta rejected a proposed remedy that would have broadly prohibited Google from favoring its Gemini system within Chrome.[70] He warned that such a restriction would impair competition:

Such a restriction would set Google apart from its competitors. … The court will not hobble Google’s competitiveness by prohibiting self-preferencing of its own GenAI technologies, when that is precisely how the emerging—and highly competitive—GenAI marketplace operates.[71]

Accordingly, extending traditional anti–Big Tech assumptions to AI markets risks counterproductive enforcement. Restricting integration strategies could weaken competitive pressure on leading AI firms and produce the opposite of the intended effect. The risk is especially pronounced where digital-platform conduct already faces significant constraints under DMA-type regulation.

V. Conclusion

Artificial intelligence is widely described as a new technological paradigm capable of reshaping market competition. Yet competition policy toward AI remains largely framed by a Big Tech–centric narrative inherited from earlier digital-market debates. Policymakers often assume that AI will reproduce the trajectory of platform markets, in which dominant firms leveraged control over gateways and data to entrench market power. The central concern is therefore under-enforcement: that intervention will arrive “too little, too late.”

This assumption explains the skepticism toward integrating AI applications into the core services of large digital ecosystems. Authorities frequently view preinstallation, default placement, and preferential integration as extensions of earlier self-preferencing concerns. The Meta AI investigation exemplifies this approach, treating the embedding of an AI assistant into a messaging platform as a conventional tying problem.

The analysis developed in this paper suggests that this analogy is incomplete. AI markets differ from earlier platform markets in economically relevant ways. Entry remains frequent, investment levels are high, and competitive leadership changes rapidly. Network effects appear weaker than in traditional platforms, data advantages are less durable, and new firms continue to attract funding and users. The success of companies such as OpenAI and Anthropic demonstrates that incumbents have not seamlessly transferred their dominance into AI markets. Even leading positions remain unstable as competing models quickly improve.

These conditions matter for policy. Where competitive dynamics are fluid and contestable, aggressive intervention may misdiagnose the source of competitive pressure. The strategies under scrutiny—partnerships, ecosystem integration, and the internal deployment of AI tools—are not uniquely exclusionary. They are standard competitive responses in a market defined by technological uncertainty and rapid innovation. Restricting such conduct may therefore reduce, rather than enhance, competitive rivalry by weakening an important constraint on leading AI firms.

A context-sensitive framework is essential in AI markets. Authorities should continue to apply established antitrust doctrine to demonstrably exclusionary conduct, but they should avoid presuming harm from integration alone. Vertical integration and product embedding often generate efficiencies, improve coordination among complementary services, and accelerate deployment of new technologies. The Meta AI case illustrates this point: the relevant question is not whether integration occurs, but whether it actually forecloses rivals that retain alternative distribution channels and the capacity to innovate.

The regulatory implications point in a different direction. Existing digital-market regimes already impose significant obligations on designated gatekeepers when AI functionality is integrated into core platform services. By contrast, standalone AI assistants and agents may fall outside these frameworks altogether, regardless of their competitive significance. The resulting asymmetry suggests that the greater risk may not be insufficient control over incumbents, but regulatory mismatch—rules aimed at yesterday’s intermediaries, rather than today’s forms of competition.

Competition policy therefore faces two symmetrical risks. Delayed intervention may allow harmful conduct to develop, but premature intervention may suppress experimentation in a rapidly evolving market. An approach focused exclusively on preventing the first risk overlooks the second.

AI markets remain uncertain, dynamic, and highly innovative. In such conditions, competition policy should prioritize evidence over analogy. Rather than assuming that AI will replicate the history of digital platforms, enforcement should recognize that vigorous competition may already be occurring—and that poorly calibrated intervention could impede it.

[1] See, e.g., Statista, Number of AI Tool Users Worldwide from 2020 to 2031 (in Millions) (2025), https://www.statista.com/outlook/tmo/artificial-intelligence/worldwide (last visited Feb. 9, 2026) (showing steady growth in global AI-tool users—from about 116 million in 2020 to roughly 350 million in 2025, with projections exceeding 1 billion by 2031).

[2] See, e.g., Org. for Econ. Co-operation & Dev. (OECD), Artificial Intelligence and Competitive Dynamics in Downstream Markets (2025), https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/11/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_c6e81d0e/ccf0624a-en.pdf; see also Amit Zac & Michal S. Gal, The Price of Advice: Experimental Evidence on the Effects of AI Recommenders (2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5637090 (finding that consumer-facing AI recommender systems measurably influence purchasing decisions).

[3] See, e.g., CB Insights, Global Number of Deals and Funding for Agentic Artificial Intelligence (AI) Startups from 2020 to 2024 (2025), https://www.statista.com/statistics/1607697/global-agentic-ai-startup-dealsand-funding (showing rapid expansion of the agentic-AI startup sector, with funding rising from about $24 million across eight deals in 2020 to roughly $3.8 billion across 162 deals in 2024); see also Deloitte, Interest in Generative Artificial Intelligence (AI) Developments in Organizations Worldwide in 2024 (2025), https://www.statista.com/statistics/1603062/interest-in-future-genai-related-developments (reporting that agentic AI ranked organizations’ most salient technological development in 2024).

[4] See, e.g., Press Release, Autorité de la Concurrence, Conversational Agents: the Autorité Starts Inquiries Ex Officio with a View to Issuing an Opinion (2026), https://www.autoritedelaconcurrence.fr/en/press-release/conversational-agents-autorite-starts-inquiries-ex-officio-view-issuing-opinion.

[5] See, e.g., Autorité de la Concurrence, supra note 4; Austl. Competition & Consumer Comm’n (ACCC), Digital Platform Services Inquiry—Final Report (2025), https://www.accc.gov.au/about-us/publications/serial-publications/digital-platform-services-inquiry-2020-25-reports/digital-platform-services-inquiry-final-report-march-2025; Autorité de la Concurrence, Opinion on the Competitive Functioning of the Generative Artificial Intelligence Sector (2024), https://www.autoritedelaconcurrence.fr/en/opinion/competitive-functioning-generative-artificial-intelligence-sector; U.K. Competition & Mkts. Auth. (CMA), AI Foundation Models—Updated Paper (2024), https://www.gov.uk/government/publications/ai-foundation-models-update-paper; Autoridade da Concorrência, Competition and Generative Artificial Intelligence (2023), https://www.concorrencia.pt/sites/default/files/documentos/Issues%20Paper%20-%20Competition%20and%20Generative%20Artificial%20Intelligence.pdf.

[6] See, e.g., Eur. Comm’n, Consultation on the First Review of the Digital Markets Act (2025), https://digital-markets-act.ec.europa.eu/consultation-first-review-digital-markets-act_en; Eur. Comm’n, Review of the Digital Markets Act—Call for Evidence, Ares(2025)6881572 (2025), https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=intcom:Ares(2025)6881572.

[7] Shirin Ghaffary & Matt Day, Amazon Sues to Stop Perplexity from Using AI Tool to Buy Stuff, Bloomberg (Nov. 4, 2025), https://www.bloomberg.com/news/articles/2025-11-04/amazon-demands-perplexity-stop-ai-agent-from-making-purchases (reporting that Amazon claims a third-party AI shopping agent failed to disclose when it purchased on behalf of users, allegedly threatening platform integrity and merchant choice); see also Perplexity Team, Bullying Is Not Innovation, Perplexity (Nov. 4, 2025), https://www.perplexity.ai/hub/blog/bullying-is-not-innovation.

[8] Surbhi Misra, Musk Says xAI to Take Legal Action Against Apple over App Store Rankings, Reuters (Aug. 12, 2025), https://www.reuters.com/sustainability/boards-policy-regulation/musk-says-xai-take-legal-action-against-apple-over-app-store-rankings-2025-08-12.

[9] See, e.g., Press Release, Autorità Garante della Concorrenza e del Mercato (AGCM), The Italian Competition Authority Launches Investigation into Meta over Abuse of Dominant Position (July 30, 2025), https://en.agcm.it/en/media/press-releases/2025/7/A576.

[10] See, e.g., Giuseppe Colangelo, Antitrust Unchained: The EU’s Case Against Self-Preferencing, 72 GRUR Int’l 538 (2023); Pablo Ibáñez Colomo, Self-Preferencing: Yet Another Epithet in Need of Limiting Principles, 43 World Competition 417 (2020).

[11] Regulation (EU) 2022/1925 of the European Parliament and of the Council of Sept. 14, 2022 on Contestable and Fair Markets in the Digital Sector and Amending Directives (EU) 2019/1937 and (EU) 2020/1828 (Digital Markets Act), 2022 O.J. (L 265) 1.

[12] See, e.g., Friso Bostoen & Jan Kramer, Is the DMA Ready for Agentic AI?, Centre on Regulation in Europe (CERRE) (2025), https://cerre.eu/publications/is-the-dma-ready-for-agentic-ai; Jan-Frederick Gohsl, Future Proofing the DMA for Agentic AI: Lessons from the AI Act, 48 World Competition 315 (2025).

[13] See, e.g., TokenRing AI, The Search Wars of 2026: ChatGPT’s Conversational Surge Challenges Google’s Decades-Long Hegemony, WRAL (2026), https://markets.financialcontent.com/wral/article/tokenring-2026-1-2-the-search-wars-of-2026-chatgpts-conversational-surge-challenges-googles-decades-long-hegemony (reporting that ChatGPT Search captured roughly 17–18% of global search queries by early 2026).

[14] The same risk may also materialize where a partnership involves two large incumbents: see, e.g., Press Release, Google, Joint Statement from Google and Apple, Google Blog (Jan.12, 2026), https://blog.google/company-news/inside-google/company-announcements/joint-statement-google-apple (announcing a multiyear collaboration under which Apple will build next-generation foundation models on Google’s Gemini models to power future Apple Intelligence features, including a more personalized Siri).

[15] See, e.g., Org. for Econ. Co-operation & Dev. (OECD), Competition in Artificial Intelligence Infrastructure (2025), https://www.oecd.org/en/publications/competition-in-artificial-intelligence-infrastructure_623d1874-en.html.

[16] OpenAI, Introducing Apps in ChatGPT and the New Apps SDK (Oct. 6, 2025), https://openai.com/index/introducing-apps-in-chatgpt?utm_source=chatgpt.com.

[17] Jaewon Kang, Walmart Partners With OpenAI to Offer Shopping on ChatGPT, Bloomberg (Oct. 14, 2025), https://www.bloomberg.com/news/articles/2025-10-14/walmart-partners-with-openai-to-offer-shopping-on-chatgpt.

[18] See, e.g., Press Release, PayPal, OpenAI and PayPal Team Up to Power Instant Checkout and Agentic Commerce in ChatGPT (2025), https://newsroom.paypal-corp.com/2025-10-28-OpenAI-and-PayPal-Team-Up-to-Power-Instant-Checkout-and-Agentic-Commerce-in-ChatGPT.

[19] See, e.g., Vidhya Srinivasan, New Tech and Tools for Retailers to Succeed in an Agentic Shopping Era, Google (Jan. 11, 2026), https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms.

[20] See, e.g., Press Release, Conselho Administrativo de Defesa Econômica (CADE), CADE to Investigate Big Techs’ Acquisitions of AI Startups (2024), https://www.gov.br/cade/en/matters/news/cade-to-investigate-big-techs2019-acquisitions-of-ai-startups; Press Release, Eur. Comm’n, Commission Launches Calls for Contributions on Competition in Virtual Worlds and Generative AI (2024), https://ec.europa.eu/commission/presscorner/detail/en/IP_24_85; Press Release, Eur. Comm’n, U.K. Competition & Mkts. Auth. (CMA), U.S. Dep’t of Just., & Fed. Trade Comm’n, Joint Statement on Competition in Generative AI Foundation Models and AI Products (2024), https://competition-policy.ec.europa.eu/about/news/joint-statement-competition-generative-ai-foundation-models-and-ai-products-2024-07-23_en; Press Release, U.K. Competition & Mkts. Auth., CMA Seeks Views on AI Partnerships and Other Arrangements (2024), https://www.gov.uk/government/news/cma-seeks-views-on-ai-partnerships-and-other-arrangements; Press Release, Fed. Trade Comm’n, FTC Launches Inquiry into Generative AI Investments and Partnerships (2024), https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-launches-inquiry-generative-ai-investments-partnerships.

[21] See, e.g., Austl. Competition & Consumer Comm’n (ACCC), supra note 5; Autorité de la Concurrence, supra note 5; Klaus Kowalski, Cristina Volpin & Zsolt Zombori, Competition in Generative AI and Virtual Worlds, Eur. Comm’n, Competition Policy Brief No. 3 (2024), https://op.europa.eu/en/publication-detail/-/publication/5530c8ca-7a1f-11ef-bbbe-01aa75ed71a1/language-en; U.K. Competition & Mkts. Auth. (CMA), supra note 5. In the literature, see, e.g., Dirk Auer & Mario Zúñiga, AI Partnerships and Competition: Damned if You Buy, Damned if You Don’t, Int’l Ctr. for L. & Econ. (2025), https://laweconcenter.org/resources/ai-partnerships-and-competition-damned-if-you-buydamned-if-you-dont; Josef Drexl & Daria Kim, AI Innovation Competition as a Discovery Procedure: The Role and Limits of Competition Law (2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5439660.

[22] See, e.g., Autorité de la Concurrence, supra note 4; Austl. Competition & Consumer Comm’n (ACCC), supra note 5; Autorité de la Concurrence, Opinion on the Competitive Functioning of the Generative Artificial Intelligence Sector, supra note 5; U.K. Competition & Mkts. Auth. (CMA), AI Foundation Models: Initial Report (2023), https://www.gov.uk/government/publications/ai-foundation-models-initial-report; Autoridade da Concorrência, supra note 5; Body of Eur. Regulators for Elec. Commc’ns (BEREC), BEREC High-Level Position on Artificial Intelligence and Virtual Worlds, BoR (24) 68 (2024).

[23] See, e.g., Margrethe Vestager, Making Artificial Intelligence Available to All—How to Avoid Big Tech’s Monopoly on AI?, Eur. Comm’n (2024), https://ec.europa.eu/commission/presscorner/detail/en/speech_24_931.

[24] Autorità Garante della Concorrenza e del Mercato (AGCM), Decision No. 31634, Case A576, Meta AI (July 22, 2025) (noting that Meta later offered the service as a standalone product at meta.ai and, at least in the United States and Canada, through a dedicated iOS and Android app).

[25] Id., ¶¶ 4 and 6.

[26] Id., ¶ 7.

[27] Id., ¶¶ 8 and 10.

[28] Id., ¶ 42.

[29] Id., ¶ 43.

[30] Id., ¶ 45.

[31] See, e.g., Case C-233/23, Alphabet Inc. & Others v. Autorità Garante della Concorrenza e del Mercato (Android Auto), EU:C:2025:110 (Feb. 25, 2025); Case C-48/22 P, Google LLC & Alphabet Inc. v. Eur. Comm’n (Google Shopping), EU:C:2024:726 (Sept. 10, 2024); Case C-252/21, Meta Platforms Inc. v. Bundeskartellamt, EU:C:2023:537 (July 4, 2023).

[32] Case T-201/04, Microsoft Corp. v. Eur. Comm’n, EU:T:2007:289 (Gen. Ct. Sept. 17, 2007).

[33] Id., ¶ 1036.

[34] Eur. Comm’n, Case AT.40099, Google Android (July 18, 2018), confirmed by Case T-604/18, Google LLC v. Eur. Comm’n, EU:T:2022:541 (Gen. Ct. Sept. 14, 2022).

[35] Eur. Comm’n, Case AT.40684, Facebook Marketplace (Nov. 14, 2024).

[36] Id., ¶ 820.

[37] Id., ¶ 821.

[38] United States et al. v. Google LLC, 747 F. Supp. 3d 1 (D.D.C. 2024).

[39] In re Google Play Store Antitrust Litig., No. 20-CV-05671-JD, 2024 WL 4438249 (N.D. Cal. Oct. 7, 2024), aff’d, 147 F.4th 917 (9th Cir. 2025), modified, 152 F.4th 1078 (9th Cir. 2025).

[40] See, e.g., Press Release, Autorità Garante della Concorrenza e del Mercato (AGCM), The Italian Competition Authority Opens Procedure for the Adoption of Interim Measures Against Meta over Abuse of a Dominant Position (2025), https://en.agcm.it/en/media/press-releases/2025/11/A576.

[41] See, e.g., Press Release, Autorità Garante della Concorrenza e del Mercato (AGCM), The Italian Competition Authority Orders Meta to Suspend the Terms Excluding Competing AI Chatbots from WhatsApp (2025), https://en.agcm.it/en/media/press-releases/2025/12/A576.

[42] See, e.g., Conselho Administrativo de Defesa Econômica (CADE), Cade Abre Inquérito Contra Meta e Aplica Medida Preventiva Suspendendo Novos Termos do WhatsApp sobre IA (2026), https://www.gov.br/cade/pt-br/assuntos/noticias/cade-abre-inquerito-contra-meta-e-aplica-medida-preventiva-suspendendo-novos-termos-do-whatsapp-sobre-ia; Press Release, Eur. Comm’n, Commission Notifies Meta of Possible Interim Measures to Reverse Exclusion of Third-Party AI Assistants from WhatsApp (2026), https://ec.europa.eu/commission/presscorner/detail/en/ip_26_310.

[43] See Autorità Garante della Concorrenza e del Mercato (AGCM), Decision No. 31775, Case A576, Meta AI ¶ 75 (Dec. 22, 2025).

[44] Id., ¶¶ 31-32.

[45] See Eur. Comm’n, supra note 6.

[46] See, e.g., Press Release, Eur. Comm’n, Commission Opens Proceedings to Assist Google in Complying with Interoperability and Online Search Data Sharing Obligations under the Digital Markets Act (2026), https://ec.europa.eu/commission/presscorner/detail/en/ip_26_202.

[47] See, e.g., Bostoen & Kramer, supra note 12.

[48] See, e.g., OECD, supra note 2; Gohsl, supra note 12.

[49] See, e.g., Eur. Comm’n, U.K. Competition & Mkts. Auth., U.S. Dep’t of Just., & Fed. Trade Comm’n, supra note 20.

[50] See, e.g., Eur. Comm’n, supra note 20; Kowalski, Volpin & Zombori, supra note 21.

[51] Autorità Garante della Concorrenza e del Mercato (AGCM), supra note 24, ¶ 36.

[52] Id.

[53] See, e.g., Anton Korinek & Jai Vipra, Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence, 40 Econ. Pol’y 227 (2025); Catherine Tucker, How Does Competition Policy Need to Change in a World of Artificial Intelligence?, 40 Oxford Rev. Econ. Pol’y 834 (2024).

[54] See, e.g., Andrei Hagiu & Julian Wright, Artificial Intelligence and Competition Policy, 103 Int’l J. Indus. Org. 103134 (2025); Korinek & Vipra, supra note 53; Zach Meyers & Marc Bourreau, A Competition Policy for Cloud and AI, Centre on Regulation in Europe (CERRE) (2025), https://cerre.eu/publications/acompetition-policy-for-cloud-and-ai; Thibault Schrepel & Alex “Sandy” Pentland, Competition Between AI Foundation Models: Dynamics and Policy Recommendations, 34 Indus. & Corp. Change 1085 (2025). See also Austl. Competition & Consumer Comm’n (ACCC), supra note 5, at 292; Autorité de la Concurrence, supra note 5, at 5.

[55] See, e.g., Kate Clark, Anthropic Raising $10 Billion at $350 Billion Value, Wall St. J. (Jan. 7, 2026), https://www.wsj.com/tech/ai/anthropic-raising-10-billion-at-350-billion-value-62af49f4.

[56] Meyers & Bourreau, supra note 54.

[57] Id.

[58] Geoffrey A. Manne & Dirk Auer, From Data Myths to Data Reality: What Generative AI Can Tell Us About Competition Policy (and Vice Versa), CPI Antitrust Chron. (Feb. 2024); see, e.g., CB Insights, Artificial Intelligence (AI) Unicorns Worldwide in 2nd Quarter 2025, by Valuation (2025), https://www.statista.com/statistics/1621613/artificial-intelligence-unicorns-worldwide (showing the AI-startup ecosystem led by ByteDance and OpenAI at roughly $300 billion valuations, followed by Stripe ($70 billion), Databricks and Anthropic ($62 billion each), and xAI ($50 billion)).

[59] See Roland Berger, Most Popular Artificial Intelligence (AI) Applications Worldwide in February 2025, by Monthly Active Users (2025), https://www.statista.com/statistics/1609163/top-ai-applications-mau-worldwide (reporting that the ByteDance-owned chatbot Doubao reached about 82 million monthly active users, with ChatGPT’s Nova Assistant at roughly 63 million and DeepSeek at about 62 million; a second group—Remini, Talkie AI, Character AI, ChatOn, Genius, and Gemini—each recorded about 28–33 million users); see also Similarweb, 2025 Generative AI Landscape: The State of Gen AI (2025), https://www.similarweb.com/corp/2025-generative-ai-landscape (finding that ChatGPT leads U.S. usage with more than 41 million monthly active users and a 33% stickiness rate, while rivals such as Perplexity, Copilot, and Gemini have smaller user bases and lower engagement).

[60] Semrush, Artificial Intelligence (AI) Chatbots Worldwide Market Share from April 2024 to March 2025 (2025), https://www.statista.com/statistics/1618020/ai-chatbots-traffic-share-ww.

[61] See Similarweb, supra note 59 (reporting that in Sept. 2025 Google received about 82 billion monthly visits worldwide, followed by YouTube at roughly 29 billion, Facebook at about 11 billion, Instagram at approximately 6.5 billion, and ChatGPT at around 6 billion).

[62] See AppMagic, Most Downloaded Generative AI Mobile Apps Worldwide as of November 27, 2025 (2025), https://www.statista.com/statistics/1554189/top-gen-ai-apps-by-downloads (showing Google Gemini with about 392 million downloads, followed by Cici (169 million), DeepSeek (158 million), Perplexity (95 million), and Grok (82 million)).

[63] See Stack Overflow, Most Used Artificial Intelligence (AI) Search and Developer Tools Among Developers Worldwide as of 2024 (2024), https://www.statista.com/statistics/1483838/ai-tools-usage-among-developers-use-worldwide (reporting GitHub Copilot usage at 44%, Google Gemini at 22%, Bing AI at 14%, and Visual Studio IntelliCode at 13.7%, with lower adoption for Claude (7.6%) and Perplexity AI (4.9%)).

[64] See Similarweb, AI Global—Global Sector Trends on Generative AI (2026), https://www.similarweb.com/corp/wp-content/uploads/2026/01/attachment-Global-AI-Tracker-6.pdf?utm_medium=social&utm_source=twit (reporting lower market shares for rivals, including DeepSeek (3.7%), Grok (3.4%), Perplexity (2.0%), Claude (2.0%), and Copilot (1.1%)).

[65] Bradley Olson, The Week Anthropic Tanked the Market and Pulled Ahead of Its Rivals, Wall St. J. (Feb. 5, 2026), https://www.wsj.com/tech/ai/the-week-anthropic-tanked-the-market-and-pulled-ahead-of-its-rivals-ef59dff1; George Hammond, Anthropic’s Breakout Moment: How Claude Won Business and Shook Markets, Fin. Times (Feb. 6, 2026), https://www.ft.com/content/a75555a6-24c3-4468-aba9-7fe12b5def31.

[66] U.K. Competition & Mkts. Auth. (CMA), supra note 22, ¶ 4.17.

[67] Semrush, supra note 60.

[68] Similarweb, supra note 64.

[69] Stack Overflow, supra note 63.

[70] United States et al. v. Google LLC, No. 20-cv-3010 (APM) (D.D.C. 2025).

[71] Id.

Abstract

As the European Directive on the Enforcement of Intellectual Property Rights (IPRED) marked the twentieth anniversary of its adoption, renewed calls have emerged for its revision, aimed at fostering a more effective application of the principle of proportionality in patent enforcement. Proponents of reform argue that injunctive relief continues to be granted in an overly automatic manner and should therefore be subject to greater restraint. To this end, it has been suggested that valuable guidance may be drawn from the U.S. legal landscape and, in particular, from the framework articulated by the U.S. Supreme Court in eBay v. MercExchange. Against this background, the paper critically examines these reform proposals, arguing that they appear to rest on the same arguments that underpinned the highly controversial regulatory proposal on standard essential patents (SEPs), and questioning the purported alignment between the European and the U.S. approaches to patent enforcement.

Abstract

Politicians need support (votes in democracies, compliance and participation in autocracies) and must pay for it through patronage, public services, and policy concessions. I model this as monopsony: the politician faces an upward-sloping supply curve of political support. Propaganda is a complement that shifts supply down by making compliance less distasteful. The politician equates the marginal cost of propaganda to the wage savings on inframarginal supporters. I prove three results. First, monopolists use more propaganda than competitive politicians because markdowns are larger and there is no freeriding on regime-level messaging. Second, coercion and propaganda are complements: coercion makes supply more inelastic (raising markdowns) and enables forced consumption of propaganda that citizens would otherwise reject. Third, the model predicts scale effects: propaganda’s returns rise with population, implying that mass-mobilization autocracies should propagandize heavily while elite autocracies rely on direct payments. Cross-country patterns are consistent with these predictions, though the scale effect is modest

Threads from ICLE scholars on trending issues for the month of February 2026.

Join ICLE in D.C. on 3/2 for "The Competition and Consumer Protection Year in Review." Experts, including @chris_mufarrige, @M_Ohlhausen, and @DanielJGilman1 will evaluate the past year of enforcement at the FTC.

Register using the link below ?? pic.twitter.com/OtoJzJJ6re

— Int'l Ctr Law & Econ (@LawEconCenter) February 18, 2026

I'm super excited for my new paper with @ATabarrok and Mark Whitmeyer:

"Chaos and Misallocation under Price Controls"

During the 1973-74 gasoline crisis, the U.S. had about a 9 percent national shortfall. But that was far from evenly spread out.

Over 90 percent of stations… pic.twitter.com/kmY0TXHRwG

— Brian Albrecht (@BrianCAlbrecht) February 12, 2026

Populists across the spectrum, left and right, hate economists. I joked its some puzzle but I think there's a simple reason. It's not about liking capitalism or something.

And the disagreement is almost always about how to reason about problems, not about values.

Populists… https://t.co/1YRKN8XclQ

— Brian Albrecht (@BrianCAlbrecht) February 11, 2026

??? #Brazil #DigitalMarkets #Regulation #Competition

Brazil’s House may vote today on fast-tracking (applying a "regime de urgência") Bill 4765/2025 –a "DMA-style" digital markets regulation. IMHO, this is *wrong* for this type of bill. A short ? explaining why: pic.twitter.com/qQcH9dTiJ1

— Mario Zúñiga (@MZunigaP) February 10, 2026

New @LawEconCenter working paper out:
“Don’t catch me (too early) if you can: a #competition policy perspective on the #integration of assistive and agentic #AI services.” ?https://t.co/LYdKByXZvm

— Giuseppe Colangelo (@GiuColangelo) February 10, 2026

Happy to share our latest @Deep_In_Network study: “Lost in Translation? #Injunctions and #Patent Enforcement in a Transatlantic Perspective.”

— Giuseppe Colangelo (@GiuColangelo) February 6, 2026

New paper: "A Price Theory of Propaganda"

Most models of propaganda are about beliefs. How do regimes manipulate what citizens think?

I take a different approach. Propaganda shifts the cost of providing political support, not the content of beliefs. pic.twitter.com/YPRENhRFq1

— Brian Albrecht (@BrianCAlbrecht) February 5, 2026

The German Bundeskartellamt just prohibited Amazon's "price control mechanisms" (Orwellian terminology) and ordered €59 million in disgorgement.

The offense?

Amazon's algorithms flagged marketplace sellers whose prices were TOO HIGH and reduced their visibility. A retailer… https://t.co/ioPUqDS21C

— Brian Albrecht (@BrianCAlbrecht) February 5, 2026

The Bundeskartellamt’s Amazon case rests on a series of analytical mistakes that recur across modern platform enforcement. Here’s a tentative list:

1.Neutrality Fallacy

The assumption that intermediation should be neutral by default, even though the value of platforms like… https://t.co/DV3Eng6s5T

— Lazar Radic (@laz_radic) February 5, 2026

Amazon fined for displaying "price-competitive" offers on its German website.

This epitomizes the deep problems that afflict competition enforcement in Europe:

– Hypothetical concerns are trumping common sense

– Enforcement of ex ante rules is further fragmenting markets

? https://t.co/g60hzt2hiL

— Dirk Auer (@AuerDirk) February 5, 2026

The Bundeskartellamt's Amazon case is built on a fundamental misunderstanding of intermediation.

It fails to consider that Amazon’s so called “discretion” might not be distortion, but the source of the platform's value. Ranking, filtering, excluding bad actors, disciplining… pic.twitter.com/3aZlqmAkR8

— Lazar Radic (@laz_radic) February 5, 2026