Toward a Simple Economic Framework for Analysis of Business Ecosystems

Abstract

This chapter develops a progression of frameworks to analyze economic organization for both single- and multi-product production, culminating in a framework for studying business ecosystems. The first stage draws on transaction cost economics, focusing on how governance affects the costs of transactions. The second incorporates transaction value economics, highlighting how interactions can create value through mechanisms such as knowledge sharing. The third extends these perspectives by integrating insights from property rights economics, along with additional insights from transaction value economics, to account for how governance influences the value of future transactions via its impact on investment and innovation, with a particular focus on the latter. This progression provides a structured, economically grounded approach for analyzing coevolution and other defining features of business ecosystems within an antitrust-relevant framework – analysis that has become especially relevant and important with the proliferation of digital business ecosystems.

Read the full piece at SSRN.

Executive Summary

Many state automobile-dealer franchise laws require manufacturers to sell vehicles through franchised dealers, even when manufacturers would prefer to sell directly to consumers. Originally enacted to regulate bargaining relationships between manufacturers and their dealer networks, these laws now function as a mandatory intermediary requirement. The result is effectively a “middleman tax” on car buyers: a legally imposed distribution layer that increases vehicle prices even where more efficient models are available.

Using an analytical framework first developed in a 2000 Goldman Sachs study and later cited by the U.S. Department of Justice, this issue brief updates the methodology with current vehicle prices, interest rates, inventory data, and dealer operating costs. The results suggest that direct distribution could generate savings of roughly $3,934 to $4,992 per vehicle. Applied to the modern average transaction price of about $50,000, this represents an efficiency gain of 7.87% to 9.87%—the practical equivalent of removing a substantial implicit surcharge from every new-car purchase.

Mandatory dealer intermediation also fragments the national market, burdens interstate commerce, and fits poorly with modern vehicle business models built around digital retailing, centralized pricing, and software-defined vehicles. The policy principle is simple: allow channel competition. States should not mandate a single distribution architecture when multiple models can compete to serve consumers. Removing the “middleman tax” would lower costs, expand consumer choice, and better align the law with the modern automobile market.

I. Introduction

For most Americans, buying a car is one of the largest purchases they will ever make. Yet in all 50 states, a web of dealer-franchise laws dictates how that purchase must occur. Legislatures enacted these statutes between the 1930s and 1970s to protect franchised dealers from coercive manufacturer practices. Today, they serve a different function: many state laws prohibit manufacturers from selling vehicles directly to consumers—even when no franchisee exists and no vertical relationship requires regulation.[1]

The core question is straightforward: why should state law force an out-of-state manufacturer to adopt a particular retail structure as the price of access to a state’s consumers? This issue brief argues that such mandates impose a large and unnecessary “middleman tax” on car buyers, fragment the national market in ways that burden interstate commerce, and fit poorly with modern direct-to-consumer and software-driven vehicle business models. Based on updated estimates, the current burden imposed by state franchise laws amounts to roughly $3,934 to $4,992 per vehicle—between 7.87% and 9.87% of the average $50,000 purchase price.[2]

This issue brief proceeds in three parts. First, it quantifies the economic costs created by this middleman tax. It builds on a conservative baseline established in a 2009 paper by the U.S. Department of Justice’s Economic Analysis Group[3] and updates that methodology using current vehicle prices, inventory data, interest rates, and compliance costs. Second, it examines the constitutional limits that the Dormant Commerce Clause places on state efforts to dictate firm organization as a condition of market access. Third, it shows that many state franchise statutes fit poorly—both textually and structurally—with the business models they now attempt to regulate. Courts and regulators often rely on expansive interpretations of provisions designed for a fundamentally different market.

At bottom, this story is about affordability and consumer welfare. Direct sales are not a luxury preference. They lower acquisition costs, reduce transaction frictions, and speed consumer access to cost-saving vehicle technologies. In a market where the average new-vehicle transaction price now exceeds $50,000[4] and affordability sits near historic lows,[5] the stakes of this debate are substantial.

II. The Origins and Evolution of Dealer-Franchise Laws

Automobile-dealer franchise laws emerged between the 1930s and 1950s, when the U.S. auto market was dominated by the “Big Three”: General Motors, Ford, and Chrysler.[6] In that environment, manufacturers were widely seen as holding overwhelming bargaining power over their franchised dealer networks. Critics argued that automakers could force dealers to accept unwanted vehicles, terminate franchise agreements without cause, or place new franchises inside an existing dealer’s territory. States responded by enacting franchise-protection statutes. By the mid-20th century, every state had adopted some form of dealer-protection law.

These laws targeted a specific problem: the balance of power within an intrabrand relationship. Legislatures sought to regulate the vertical relationship between a manufacturer and its own franchised dealers. The entire statutory framework assumed the existence of that relationship and focused on the bargaining power within it.[7]

The modern automobile market looks very different. The Big Three’s combined market share has fallen from roughly 90% in the 1950s to about 40% today.[8] Interbrand competition now defines the market, with dozens of manufacturers competing for consumers. Dealers have also changed. What were once largely small, family-owned businesses have evolved into large, sophisticated enterprises, often owned by multi-state groups or publicly traded companies. As of 2024, the United States had about 16,957 franchised light-vehicle dealerships.[9] Currently, about 150 dealerships control roughly 30% of industry revenue, and some estimates project that figure will reach 50% by 2050.[10]

This shift highlights a categorical distinction. Regulating an existing vertical relationship between a manufacturer and its franchisees differs fundamentally from prohibiting an alternative distribution model where no franchisee exists. New entrants such as Tesla, Lucid, Rivian, Scout, and others operate without franchised dealers. There is no intrabrand relationship to regulate—a point recognized by at least one state supreme court.[11] Applying franchise-protection statutes to these firms therefore reflects a category error. It conflates the regulation of an existing franchise relationship with the prohibition of a competing business model.

The persistence of these laws, despite the disappearance of the market conditions that produced them, illustrates a classic pattern of regulatory capture.[12] Automobile dealers—the organized interest group that benefits from the statutes—have used the legislative process to preserve economic rents, while spreading the resulting costs across consumers.

III. The Economic Cost of the ‘Middleman Tax’

Dealer-franchise laws operate, in economic terms, as a structural restraint on distribution. They require a two-tier channel—manufacturer to dealer to consumer—even where a one-tier channel—manufacturer to consumer—would be feasible and potentially more efficient. This requirement effectively imposes what this issue brief refers to as a “middleman tax” on vehicle purchases.

The resulting costs arise from several sources. Maintaining dealer inventories ties up capital and slows inventory turnover. Sales commissions and related transaction overhead increase the cost of each sale. Thousands of retail locations duplicate facilities, staffing, and administrative functions that a more centralized distribution system could avoid. Consumers also bear bargaining and search costs that accompany dealership negotiations. Finally, the make-to-stock distribution model weakens the link between production and consumer preferences, producing misallocation costs that appear as rebates, incentives, and markdowns on slow-moving inventory.[13] Although not captured in the foregoing analysis as an economic cost, the dealership model also imposes substantial transaction costs on consumers. The average vehicle purchase required roughly 13 hours of buyer time in 2025, including research, dealer visits, and negotiation.[14]

The sections that follow quantify these effects and place them in a broader economic context. Section III.A summarizes the baseline estimates developed by the U.S. Department of Justice’s Economic Analysis Group. Section III.B updates those estimates using current vehicle prices, interest rates, and dealer operating costs, producing a modern estimate of the middleman tax imposed by dealer mandates. Section III.C then examines the dynamic effects of these rules, showing how dealer mandates can dampen manufacturing investment and constrain innovation in software-defined vehicles and integrated digital-service platforms.

A. The DOJ Baseline Estimate

The most rigorous government analysis of these costs appears in a 2009 paper by the U.S. Department of Justice’s (DOJ) Economic Analysis Group, authored by Gerald R. Bodisch.[15] The Bodisch paper established several conservative benchmarks for understanding the costs of the traditional dealer-based distribution system.

First, Bodisch estimated that the total cost of automobile distribution in the United States averaged as much as 30% of the final vehicle price. These costs split roughly evenly between manufacturer-related expenses—such as advertising, rebates, and subsidized financing—and dealer-related expenses, including inventory financing, insurance, advertising, and sales commissions.

Second, the paper documented the scale of capital tied up in the industry’s “make-to-stock” distribution model. At the end of 2008, the nation’s roughly 20,700 franchised dealerships collectively held about $100 billion in new-vehicle inventory.[16] Bodisch estimated that carrying this inventory imposed annual costs of roughly $890 million.

Third—and most relevant for present purposes—the paper relied on a 2000 Goldman Sachs analysis of the potential savings from shifting to a direct-sales, build-to-order distribution model. Using an average vehicle price of $26,000, Goldman Sachs estimated total savings of $2,225 per vehicle, or roughly 8.6% of manufacturer’s suggested retail price (MSRP).[17] These savings fell across five categories, illustrated in Table I.

TABLE 1: Estimated Per-Vehicle Savings from Shift to Direct-Sales Model

SOURCE: Goldman Sachs (2000)

TABLE II: Range of Savings Across Multiple Price Points[18]

Goldman Sachs also identified roughly $1,000 in additional potential savings per vehicle from improvements in product development, manufacturing flexibility, and procurement. The DOJ paper attributed most of the benefits, however, to efficiencies created by shorter order-to-delivery cycles.[19]

To illustrate that these projections were not purely theoretical, the paper pointed to the Chevrolet Celta program in Brazil. There, General Motors implemented a build-to-order, internet-direct model that produced consumer prices roughly 6% lower, while selling more than 700,000 vehicles.[20]

B. Updating the DOJ Baseline

The Goldman Sachs estimates relied on data from 2000 and were reported by the DOJ in 2009. Many of the original inputs—vehicle prices, floorplan interest rates, facility costs, and inventory carrying charges—are now outdated. Importantly, most changes since 2000 have made the physical dealer infrastructure more expensive to maintain, not less.

This issue brief therefore applies the same component framework to current market conditions. The resulting figures should be understood as conservative. A fresh bottom-up reconstruction using current dealer financial data would likely produce equal or larger estimates for most components. Appendix A provides the detailed methodology, underlying data sources, and academic cross-references.

Several key inputs have shifted since the original analysis. First, the average transaction price (ATP) of a new vehicle has roughly doubled, surpassing $50,000 for the first time in September 2025.[21] Second, the interest-rate environment has tightened considerably. Dealer floorplan financing—typically structured as variable-rate facilities priced at the Secured Overnight Financing Rate (SOFR) plus 200–400 basis points—now carries effective rates of roughly 6% to 9% for well-qualified dealers, substantially higher than early-2000s levels.[22]

Other structural conditions have changed far less. Days’ supply for new vehicles still falls in the 60-90 day range identified by Bodisch, suggesting that the core inefficiencies of the make-to-stock distribution model persist despite two decades of digital transformation.[23] Dealership operating costs also remain substantial. According to the National Automobile Dealers Association (NADA), the average franchised dealership generated $73.3 million in total sales in 2024, while earning net margins of roughly 1% to 2%.[24] Those margins imply non-vehicle-acquisition operating expenses—including payroll, facilities, advertising, insurance, and floorplan interest—of about $500,000 or more per month.

Applying the Goldman Sachs component ratios to a $50,000 ATP produces the updated estimates illustrated in Table III.

TABLE III: Updated Savings Estimates for Shift to Direct-Sales Model

The updated analysis suggests that direct distribution could generate savings of roughly $3,934 to $4,992 per vehicle—up to a 9.87% efficiency gain when applied to the modern ATP.[25]

These estimates remain conservative. Elevated interest rates, with floorplan borrowing costs around 6% or higher, increase inventory carrying costs more than simple ratio scaling would suggest. Bottom-up reconstructions of individual components confirm that the ratio-based estimate likely understates the true savings. Appendix A provides the full component-level analysis, methodology, data sources, and academic cross-references.

C. Investment and Innovation Costs of Dealer Mandates

Dealer-franchise laws impose more than static distribution costs. They also create dynamic costs for manufacturing investment and innovation. Lower effective consumer prices—made possible by direct distribution—expand the addressable market for new entrants. As prices fall, more consumers can afford to purchase vehicles, increasing unit demand and improving capacity utilization at domestic production facilities. Higher volumes, in turn, make it easier for firms to justify large-scale manufacturing investments in the United States. While any estimate of incremental production is necessarily counterfactual, the underlying economic logic is straightforward: lower prices expand demand.

The innovation channel is equally important. Modern vehicles are increasingly defined by software. Over-the-air (OTA) updates, subscription-based features, and integrated digital-service ecosystems now shape how manufacturers design and deliver vehicles. These business models do not fit neatly within the mid-20th-century dealership framework—if they fit at all—which assumed that a “sale” occurred as a discrete transaction at a physical location, followed by periodic warranty or maintenance visits to a local dealer.

The tension arises because software-defined vehicles depend on an ongoing digital relationship between the manufacturer and the vehicle owner. Over-the-air (OTA) software updates allow manufacturers to remotely update vehicle systems, add features, or correct defects without requiring a visit to a dealership. These updates are delivered directly to vehicles through wireless networks, much like updates to smartphones or other connected devices.

By eliminating the need for in-person service visits, OTA updates reduce recall costs and enable continuous improvement of vehicle functionality. At the same time, they weaken the traditional dealership role as the primary intermediary for maintenance and customer contact. Dealerships derive substantial revenue from service visits and warranty work, so this shift creates a structural conflict between the software-centric design of modern vehicles and a distribution system built around periodic dealership interaction.

Manufacturers operating under a direct-sales model can instead iterate continuously on the customer experience. They can integrate vehicle purchase, software management, service scheduling, and charging infrastructure into a unified digital platform. Dealer-franchise laws often block this approach for new entrants. In many states, those statutes effectively require manufacturers to route vehicle sales through a franchised dealer network—even when no such network exists.

IV. The Dormant Commerce Clause and Dealer-Franchise Laws

The Dormant Commerce Clause is the constitutional principle that limits states’ ability to interfere with interstate commerce. Although the U.S. Constitution grants Congress the power to regulate commerce among the states, the Supreme Court has long interpreted that grant to carry a negative implication: states may not enact laws that discriminate against interstate commerce or impose burdens that outweigh their local benefits. In practice, the doctrine prevents individual states from erecting protectionist barriers or regulating commerce in ways that disrupt the national market.

State dealer-franchise laws implicate this doctrine because they dictate the business structure an out-of-state manufacturer must adopt to access a state’s consumers. These statutes require manufacturers to distribute vehicles through franchised dealers, rather than sell directly to consumers. In effect, they condition market access on adopting a state-mandated distribution model.

Direct-sales bans are vulnerable to challenge on two grounds. First, they discriminate in practical effect against out-of-state manufacturers that prefer distribution models that do not rely on in-state franchised dealers. Although these statutes appear facially neutral, they systematically advantage the incumbent dealer network—a politically powerful in-state interest group—while disadvantaging out-of-state manufacturers that seek to compete through alternative channels.

Second, even if a court determines that a direct-sales ban is not discriminatory, the statute must still survive the balancing test the Supreme Court articulated in Pike v. Bruce Church, Inc. Under that test, a state law is invalid if “the burden imposed on such commerce is clearly excessive in relation to the putative local benefits.”[26] On the burden side, the analysis is straightforward. Dealer-franchise laws impose billions of dollars in aggregate costs on consumers, fragment the national market into 50 regulatory regimes, and block the entry of competing business models. On the benefit side, the asserted local interests—consumer protection, local employment, and service infrastructure—are either pretextual or achievable through less restrictive means.[27]

Arguments that direct-sales bans protect public health or consumer welfare have repeatedly been rejected by legal scholars, economists, and staff of the Federal Trade Commission (FTC).[28] The central problem is pretext. Dealer-franchise laws were enacted to protect dealers from manufacturer coercion, not to protect consumers from manufacturers. Recasting those statutes as consumer-protection measures after the fact does not change their legislative purpose or practical effect.

Even accepting the asserted benefits at face value, categorical bans on direct sales are not narrowly tailored to achieve them. States possess a range of less restrictive alternatives. They can license direct-sales facilities, impose disclosure requirements, enforce lemon laws against all sellers regardless of distribution model, and require service-availability guarantees as a condition of market access. None of these tools requires banning an entire distribution channel.

An additional—and often overlooked—dimension of the Dormant Commerce Clause challenge involves extraterritoriality. When individual states dictate the distribution structure manufacturers must adopt, they effectively force firms to organize their national operations around the most restrictive state’s rules. Manufacturers cannot realistically maintain one distribution system for permissive states and another for restrictive states without incurring large duplicative costs. The result is that a handful of restrictive states determine the distribution architecture for the entire national market—precisely the kind of extraterritorial regulation the Commerce Clause was designed to prevent.[29]

Courts and policymakers have begun to grapple with these tensions. In Massachusetts, the Supreme Judicial Court held that the state’s dealer association lacked standing to challenge Tesla’s direct-sales operations because Tesla had no affiliated dealer network. The court recognized that Tesla’s business model presented no franchise relationship to regulate.[30] In Illinois, an appellate court ruled in 2024 that existing law does not bar manufacturers from obtaining dealer licenses to conduct direct sales.[31]

Manufacturers continue to test these legal boundaries. Lucid has filed challenges in multiple states.[32] Rivian has backed reform efforts in Washington state, including a now-abandoned 2026 ballot initiative to authorize direct-to-consumer EV sales.[33] Rivian has also filed suit in Ohio challenging a state direct-sales ban that permits Tesla to sell vehicles directly while excluding newer entrants.[34]

Taken together, these developments suggest that courts and legislatures[35] are beginning to recognize a key distinction: regulating existing franchise relationships differs fundamentally from prohibiting alternative business models altogether.

V. Statutory Misfit in the Modern Vehicle Market

Even before reaching constitutional questions, many state franchise statutes fit poorly with modern vehicle business models. Courts and regulators often apply these laws to direct-to-consumer manufacturers through expansive interpretations of provisions designed for a fundamentally different commercial relationship.

Most franchise statutes revolve around the manufacturer-dealer relationship. Their operative provisions govern franchise termination, territorial allocations, facility requirements, and other features of an ongoing vertical business arrangement. Direct-to-consumer entrants have none of these elements. They operate without franchisees, franchise agreements, or territorial allocations. Applying franchise statutes to such firms does not advance the statutes’ original purpose. Instead, it stretches the laws beyond their text and structure to reach conduct the legislature never contemplated.[36]

The modern automotive market now includes business models that diverge sharply from the franchised-dealer archetype. Manufacturers increasingly rely on online ordering with centralized, non-negotiable pricing; agency sales models in which the “dealer” functions primarily as a delivery agent, rather than a reseller; remote delivery and mobile-service routing; software-defined vehicle features delivered through over-the-air updates; vertically integrated charging and service networks; and hybrid retail models in which showrooms exist but do not operate as statutory “dealers.” These approaches challenge the core assumptions of franchise law, which presupposed a manufacturer producing a finished vehicle, shipping it to a dealer’s lot, and allowing the dealer to negotiate the retail price with the consumer.

State regulatory approaches to these models vary widely. In practice, state regimes generally fall into four categories: outright bans on direct manufacturer sales; bans with limited electric-vehicle carve-outs; capped regimes that permit direct sales subject to limits on the number of stores or vehicles sold; and permissive regimes that allow direct sales with few restrictions. This patchwork fragments the national market and complicates manufacturers’ efforts to operate under a consistent distribution model.

In many restrictive states, regulators stretch the statutory definitions of “dealer” and “franchise” beyond recognition to capture manufacturers that have never operated through a franchise model. This interpretive expansion is not merely technical. It determines whether new competitors can reach millions of consumers.

It also bears noting that federal law already addresses many of the concerns that state franchise statutes claim to address. Federal dealer-protection statutes govern manufacturer-dealer relationships nationwide.[37] The FTC has also attempted to regulate dealer practices through rulemaking. Although a federal court vacated the agency’s proposed CARS Rule on procedural grounds,[38] the rule illustrates that federal regulators can pursue consumer-protection objectives through less restrictive tools than blanket bans on direct distribution. The presence of these federal mechanisms weakens the argument that state-level distribution bans are necessary to protect consumers.

VI. Conclusion

State automobile-dealer franchise laws impose what this issue brief calls a “middleman tax” on American car buyers. Applying the DOJ’s methodology with current market data, the analysis here estimates that mandatory dealer intermediation adds roughly $4,687 to the cost of a new vehicle. That figure rises further when dealer add-ons, bargaining rents, and administrative fees are included. In a market where the average transaction price exceeds $50,000 and affordability sits near historic lows, the issue is not abstract. It directly affects millions of consumers.

These estimates are conservative. Even using the lowest plausible inputs for each component, the floor estimate remains about $3,934 per vehicle, or 7.87% of the average transaction price. The increase in the average transaction price—from $26,000 in 2000 to about $50,000 today—largely tracks cumulative inflation. That means the structural cost of mandatory intermediation remains at least as large in real terms today as when the DOJ first documented the problem more than 15 years ago. Current market conditions likely increase the burden further. Higher interest rates, rising commercial real-estate costs, and greater vehicle complexity all raise the cost of maintaining a large brick-and-mortar dealer network. Bottom-up reconstructions using current dealer financial data consistently match or exceed the ratio-scaled estimates. A realistic estimate places the cost closer to $4,992 per vehicle, even before accounting for bargaining frictions, dealer add-ons, and consumer time costs documented in the academic literature.

Dealer franchise laws also impose broader economic costs. They fragment the national market by forcing manufacturers to organize distribution around the most restrictive state rules. They burden interstate commerce by conditioning market access on a state-mandated business structure. And they fit poorly with modern vehicle business models, extending franchise-protection statutes designed for mid-20th-century manufacturer-dealer relationships to firms that have never operated through a franchise system.

The principle guiding reform should be simple: allow channel competition. States should not mandate a single distribution architecture when multiple models can compete to serve consumers.

Several reform paths are available. The most comprehensive option is repeal of direct-sales bans, allowing manufacturers to choose any lawful distribution model. Where repeal proves politically infeasible, states can replace categorical bans with licensing and disclosure requirements that ensure consumer protection without prohibiting direct sales. States can also require service availability, bonding, or escrow provisions as conditions of market access. Any reform should include sunset clauses and periodic review requirements to reduce the risk of regulatory recapture by incumbent interests.[39]

The federal government also has a role to play. The DOJ’s Anticompetitive Regulations Task Force could issue a formal report updating the 2009 Bodisch analysis with current market data and documenting the competitive harms of state direct-sales bans. The DOJ could also file statements of interest or amicus curiae briefs in ongoing litigation challenging these laws. In cases where state statutes conflict with the Dormant Commerce Clause or federal policy, the federal government may also consider direct legal challenges.

Congress also has a role to play. For example, it could structure relevant federal incentive programs to encourage states to adopt more forward-thinking direct-sales policies.

The case for reform is strong. Protecting an incumbent distribution channel is not the same as protecting consumers. Allowing manufacturers to compete through different distribution models would lower costs, expand consumer choice, and better align the law with the realities of the modern automobile market.

APPENDIX A: Methodology and Data Sources for the Reconstructed Goldman Sachs Distribution Savings Decomposition

I. Methodology and Baseline Assumptions

This appendix documents the methodology, data inputs, and academic cross-references underlying the updated distribution-cost estimates presented in Section III.B of the issue brief. The analytical framework originates in a 2000 Goldman Sachs equity-research report. That report estimated that shifting from the franchised-dealer distribution model to a direct-sales, build-to-order system could reduce distribution costs by roughly $2,225 per vehicle, or 8.56% of the then-average transaction price of $26,000.[40] The DOJ’s Economic Analysis Group later adopted this framework in a 2009 discussion paper to demonstrate that state dealer-franchise laws impose measurable efficiency costs on automobile distribution.[41]

The Goldman Sachs analysis relied on economic conditions that differ substantially from today’s market. Many of the original inputs—including floorplan financing rates, facility costs, and inventory-carrying charges—reflect early-2000s conditions. Over the past quarter-century, rising vehicle prices, fluctuating interest rates,[42] and higher commercial real-estate costs[43] have altered the cost structure of dealer-based distribution. These changes generally increase, rather than reduce, the cost of maintaining a large physical dealership network.

The reconstruction presented here applies the original percentage decomposition to updated market data from 2025–2026. This approach provides a conservative estimate. A fully bottom-up reconstruction using current dealer financial data would likely generate equal or larger estimates for most cost components.

The analysis uses an average transaction price (ATP) of $50,000 as the primary baseline. This figure reflects the September 2025 record of $50,080 reported by Cox Automotive’s Kelley Blue Book.[44] Each component estimate presented below draws on current industry data and, where available, cross-references the academic literature summarized in the accompanying literature review.

II. Original Component Decomposition

The 2000 Goldman Sachs report decomposed potential distribution-cost savings into five categories, each expressed as a percentage of the average vehicle price.[45] The DOJ’s Economic Analysis Group adopted these estimates in its 2009 analysis, noting that total distribution costs could reach as much as 30% of the final vehicle price, split roughly evenly between manufacturer-side and dealer-side expenses.[46]

APPENDIX TABLE I: Goldman Sachs 2000 Baseline

SOURCE: Goldman Sachs (2000)

The Goldman Sachs report also identified roughly $1,000 per vehicle in additional savings from secondary sources, including modular vehicle architecture and tighter supplier integration. Those effects are not included in the five-component decomposition reported above.[47]

III. Updated Macroeconomic and Industry Inputs

The reconstruction requires updating the key macroeconomic and industry variables that drive each savings component. The table below summarizes the baseline and updated values:

Updating the original framework requires revising the macroeconomic and industry variables that drive each savings component. Table II below summarizes the baseline values used in the 2000/2009 analysis and the updated inputs used in the reconstruction.

APPENDIX TABLE II: Updated Economic Inputs (2025–2026)

[48] [49] [50] [51] [52] [53] [54] [55] [56]

A. Note on ATP Selection

This analysis uses $50,000 as a round-number proxy for the modern average transaction price (ATP). Cox Automotive reported an ATP of $50,080 in September 2025,[57] while the December 2025 figure was approximately $47,104 for all light vehicles.[58] MoneyGeek projects an average ATP of $48,841 for 2026.[59] The $50,000 baseline therefore sits near the upper end of recent estimates. Sensitivity to ATP assumptions is addressed in Section VI below.

IV. Component-by-Component Reconstruction

This section reconstructs each component of the Goldman Sachs distribution-savings framework using updated market data. Subsection A examines savings from improved matching of production to consumer demand under a build-to-order or direct-sales model. Subsection B estimates reductions in inventory-carrying costs resulting from shorter supply chains and lower days’ supply. Subsection C analyzes potential savings from reducing dealership density and the fixed facility overhead associated with large retail networks. Subsection D evaluates reductions in sales commissions and personnel costs that arise when negotiated dealership sales processes are replaced with centralized, non-negotiable pricing. Subsection E estimates potential logistics savings from streamlined hub-to-consumer delivery and reduced dealer-network routing, and concludes with an empirical cross-check using Tesla’s operating experience as a direct-sales manufacturer. Together, these components provide a conservative reconstruction of the distribution-cost savings discussed in Section III.B of the issue brief.

A. Component 1: Improved Matching of Supply with Consumer Demand

Original Ratio: 3.20% of vehicle price.[60]

Mechanism: Under the make-to-stock model, dealers must clear unpopular configurations through incentives and discounts. A build-to-order (BTO) or direct-sales model reduces the need for such price-clearing mechanisms by aligning production with confirmed consumer preferences. The Goldman Sachs framework attributed the largest share of projected savings to this channel.[61]

Calculation:

S₁ = ATP × 0.032

S₁ = $50,000 × 0.032 = $1,600

1. Supporting data

Manufacturer incentive spending reached roughly 7.5% of ATP (about $3,750 per vehicle) by late 2025,[62] indicating that the costs of mismatched inventory remain substantial. Academic research corroborates that search and matching frictions generate meaningful price distortions. Charles Murry and Yiyi Zhou estimate that search frictions create markups of roughly $333 per vehicle in markets where dealers collocate.[63] That figure captures only consumer-side search costs, not the manufacturer-side incentives required to move mismatched inventory. Florian Zettelmeyer, Fiona Scott Morton, and Jorge Silva-Risso find that internet-based information reduces transaction prices by roughly 1.5%–2%, largely through improved consumer matching and bargaining leverage.[64] These results support the core mechanism in the Goldman Sachs framework.

2. Assessment

The 3.2% ratio remains plausible and may even be conservative under current market conditions. In 2000, incentive spending represented a smaller share of vehicle prices. By 2025, manufacturer incentives reached about 7.5% of ATP (roughly $3,750 per vehicle),[65] far exceeding the $832 the original model attributed to supply-demand mismatching at a $26,000 ATP. This pattern suggests that inefficiencies associated with make-to-stock distribution have grown relative to the original estimate. The $1,600 figure should therefore be understood as the maximum savings achievable under full BTO adoption. Even a partial shift toward build-to-order distribution could capture a meaningful share of these gains.

B. Component 2: Lower Inventory Carrying Costs

Original Ratio: 2.21% of vehicle price.[66]

Mechanism: The franchised-dealer model requires maintaining roughly 60–90 days’ supply of finished vehicles on dealer lots, generating substantial floorplan interest and operational holding costs. Direct-sales models can reduce this inventory buffer by aligning production more closely with confirmed consumer demand.[67]

Ratio-Based Calculation:

S₂ (ratio) = ATP × 0.0221

S₂ = $50,000 × 0.0221 = $1,105

Bottom-Up Reconstruction: Because the full Goldman Sachs report is not publicly available, this appendix also reconstructs inventory savings using current data inputs. Inventory-carrying costs include both an interest component and a noninterest operational component:

S₂ (reconstructed) = (ATP × r × Δt / 365) + (Δt × Cₒₚ)

Where:

• r ≈ 6% (approximate floorplan rate, conservative relative to current prime-plus dealer financing);[68]
• Δt = 45 days (estimated reduction from a 73-day legacy supply to roughly a 28-day direct-to-consumer target);[69] and
• Cₒₚ = $15 per day (noninterest holding costs, including insurance, storage, and lot maintenance).[70]

S₂ (reconstructed) = ($50,000 × 0.06 × 45 / 365) + (45 × $15)

= $370 + $675

= $1,045

The bottom-up reconstruction yields $1,045 (2.09% of ATP), slightly below the $1,105 (2.21%) produced by ratio scaling. The true value likely falls within a range of roughly $1,045–$1,105 per vehicle, or about 2.09%–2.21% of the vehicle price. Recent industry data show net floorplan expense per unit rising 35% to $487 in the first quarter of 2025, reflecting the upward pressure of higher interest rates. The 6% floorplan rate assumed here is conservative. With prime rates between roughly 6.5% and 7.5% over the past year,[71] actual dealer floorplan rates often approach or exceed 7%, which would increase the reconstructed estimate.

1. Academic cross-check

Empirical research supports the importance of the inventory channel. Gérard P. Cachon and Marcelo Olivares show that distribution-network design strongly influences finished-goods inventory levels and that inter-dealer competition encourages excess stocking.[72] Their findings support the assumption that a direct-sales model with fewer distribution points would reduce days’ supply. Adam Copeland, Wendy E. Dunn, and George Hall document systematic within-model-year price declines under the build-to-stock model, as dealers discount aging inventory to clear lots.[73]

2. Assessment

The bottom-up reconstruction yields $1,045 (2.09% of ATP), modestly below the ratio-scaled estimate of $1,105 (2.21%). Both reconstruction inputs are deliberately conservative. The assumed 6% floorplan rate sits below current prime-based dealer financing rates, which often fall in the 7%–10% range. The $15-per-day noninterest holding cost also represents a lower-bound estimate derived from aggregate dealership operating data. Adjusting either input to reflect current market conditions would increase the reconstructed value. For consistency with the original Goldman Sachs framework, this brief uses the ratio-based figure ($1,105) in the summary table. That figure falls comfortably within the range supported by current data and likely understates inventory-carrying costs in the present interest-rate environment.

C. Component 3: Reduction in Dealership Density and Facility Overhead

Original Ratio: 1.49% of vehicle price.[74]

Mechanism: Physical dealerships impose large fixed costs, including facility leases or mortgages, insurance, utilities, maintenance, and administrative personnel. A direct-sales model replaces dispersed suburban dealership lots with centralized delivery hubs, service centers, and smaller retail galleries, reducing the need for a dense network of full-service retail locations.

Ratio-Based Calculation:

S₃ (ratio) = ATP × 0.0149

S₃ = $50,000 × 0.0149 = $745

Bottom-Up Reconstruction: This component can also be reconstructed using current dealership-level data. As of 2024, the United States had approximately 16,957 franchised light-vehicle dealerships.[75] NADA reports that the average franchised dealership generated $73.3 million in total sales in 2024 on net margins of approximately 1–2 percent, with cost of goods sold (vehicle acquisition) representing roughly 90 percent of revenue. This implies non-vehicle-acquisition operating expenses on the order of $500,000 or more per month per dealership, encompassing payroll, facility costs, advertising, insurance, and floorplan interest.

Note on Cost Allocation: These operating expenses include both facility-related overhead (leases or mortgages, utilities, property insurance, taxes, building maintenance, and management) and personnel-related costs (sales commissions, salaries, and benefits). Because the Goldman Sachs framework treats commissions separately, as in Component 4, this analysis allocates only a portion of operating expenses to facility overhead to avoid double counting.

No single authoritative source reports the facility-related share of dealership operating costs. This appendix estimates facility-related overhead at roughly $100,000–$125,000 per month per dealership, covering lease or mortgage payments, utilities, insurance, property taxes, and facility maintenance. This assumption is conservative. A typical multiacre suburban dealership with showroom space, service bays, and large vehicle lots likely incurs facility costs at or above this range, particularly given commercial real-estate appreciation since 2000. Even the lower bound of $100,000 represents only about 20% of total estimated monthly operating expenses.

Calculation (Facility-Related Overhead, Assuming a 75% Network Reduction):

Total annual facility overhead:

16,957 dealerships × $112,500 per month (midpoint) × 12 months

= $22.9 billion per year

Per vehicle sold (14.9 million seasonally adjusted annual rate, or SAAR)[76]:

$22.9 billion ÷ 14.9 million = $1,536 per vehicle

Assuming a 75% reduction in the physical retail network under a direct-sales model:

S₃ (reconstructed) = $1,536 × 0.75 = $1,152 (2.3% of ATP)

For comparison, if facility-related overhead were as low as $50,000 per month—capturing only lease or mortgage costs and excluding other facility expenses—the floor estimate would be:

16,957 × $50,000 × 12 ÷ 14.9 million × 0.75

= $512 per vehicle (1.02% of ATP)

The bottom-up reconstruction therefore produces a range of $512–$1,152 per vehicle. The broader facility-overhead estimate ($1,152, or 2.3%) exceeds the ratio-scaled Goldman Sachs estimate of $745 (1.49%). The Goldman Sachs figure falls near the midpoint of the reconstructed range, suggesting that it captured more than bare lease costs but less than the full allocation of facility overhead.

A 75% reduction in the retail network is itself conservative relative to observed direct-sales models. Tesla serves U.S. sales volumes comparable to those of a mid-size franchised manufacturer through roughly 276 retail locations, whereas a similarly sized franchised brand typically operates between about 1,200 and 4,000 dealerships. That comparison implies network reductions of roughly 77%–93% on a per-brand basis. This appendix uses the lower figure to reflect that a mature direct-sales system would still require delivery hubs, service centers, and test-drive locations.

1. Academic cross-check

Empirical research on selling, general, and administrative (SG&A) costs shows that these expenses often exhibit “stickiness,” adjusting slowly when demand declines.[77] This literature suggests that overhead embedded in a mandated intermediary structure can persist even when more efficient configurations become available.[78] Scholarship on franchise law similarly characterizes dealership infrastructure as an institutional artifact: when intermediation is mandatory and entry and exit are constrained, the physical retail network persists even if market conditions would otherwise favor consolidation or alternative distribution models.

2. Assessment

The bottom-up analysis brackets the Goldman Sachs estimate. A conservative lease-only assumption supports savings of about $512 per vehicle (1.0%), while a broader facility-overhead allocation supports roughly $1,152 per vehicle (2.3%). The original Goldman Sachs ratio of 1.49% ($745) falls comfortably within this range and serves as a reasonable midpoint estimate. This appendix therefore adopts the Goldman Sachs ratio as the lower-bound estimate. Changes in underlying cost drivers since 2000 reinforce that conclusion. Commercial real-estate prices, property insurance, and utility costs have all risen substantially, increasing the per-unit cost of maintaining large suburban dealership facilities relative to the environment in which the original estimate was developed.

D. Component 4: Lower Sales Commissions and Personnel Expenses

Original Ratio: 1.47% of vehicle price.[79]

Mechanism: The franchised-dealer sales process relies on multiple layers of commissioned and salaried personnel: front-line salespeople, sales managers who approve and negotiate deals, finance-and-insurance (F&I) managers who structure financing and sell add-on products, and administrative staff who support the transaction. A direct-sales model replaces this structure with salaried product specialists operating under centralized, non-negotiable pricing. That shift reduces the multi-layered personnel overhead associated with commission-based bargaining.

Ratio-Based Calculation:

S₄ (ratio) = ATP × 0.0147

S₄ = $50,000 × 0.0147 = $735

Partial Bottom-Up Reconstruction: Average gross profit per new vehicle retailed was $2,247 in 2024. Industry-standard commission rates range from about 20%–25% of front-end gross profit.[80] Applying that range to the reported gross profit yields a direct per-vehicle commission of roughly $449–$562,[81] establishing a verifiable floor for the commission component (about 0.90%–1.12% of ATP).

Front-line commissions represent only part of the sales-department personnel cost that a direct-sales model would restructure. The remainder of the Goldman Sachs estimate—about $286 per vehicle (the difference between $735 and the lower-bound commission of $449)—likely reflects additional personnel layers involved in the traditional dealership sales process, including:

• Sales-manager compensation (deal negotiation, desk approval, and pricing authority)
• F&I-manager compensation (financing structuring, product sales, and compliance)
• Administrative and desk staff supporting the multi-step transaction process
• Training and onboarding costs for a high-turnover commissioned sales force

No single source reports per-vehicle figures for each of these components. A conservative allocation nonetheless aligns with the Goldman Sachs total: $449 in direct commissions plus roughly $286 in supporting sales-department personnel costs equals $735 (1.47% of ATP). This estimate implies that for every dollar paid in front-line commission, about $0.64 supports the surrounding sales infrastructure—a plausible ratio given the management-intensive and negotiation-driven nature of the franchised-dealer sales process.

1. Academic cross-check

Academic research documents the welfare costs associated with commission-driven bargaining models. Ambarish Chandra, Sumeet Gulati, and James Sallee find that negotiated pricing disproportionately disadvantages consumers with higher bargaining disutility, including many older buyers and women.[82] In effect, bargaining can function as a regressive transaction tax paid through time, stress, and higher expected prices. Meghan R. Busse, Jorge Silva-Risso, and Florian Zettelmeyer show that consumer-facing rebates pass through more directly to buyers than dealer-targeted incentives, suggesting that intermediaries capture surplus through opaque pricing channels.[83] Andreas Grunewald et al. estimate that eliminating dealer discretion in loan pricing alone would generate significant consumer-surplus gains—an additional cost channel beyond the commission savings estimated here.[84]

2. Assessment

The direct commission estimate of $449–$562 per vehicle provides a well-supported lower bound. The Goldman Sachs ratio of 1.47% ($735) plausibly captures the full sales-department personnel structure, although the portion above the direct commission (about $286) reflects supporting infrastructure costs that cannot be independently verified with the same precision. Some portion of this estimate may overlap with personnel costs embedded in dealership overhead (Component 3). That overlap is likely minimal because the bottom-up reconstruction in Component 3 relies primarily on facility and occupancy costs, rather than staffing figures.

The $735 estimate therefore remains a reasonable midpoint for the personnel-cost component. Importantly, it captures only the direct payroll effects of restructuring the sales process. It does not include broader welfare gains from eliminating bargaining frictions—such as reduced consumer time costs, lower price dispersion, and improved transparency—which the academic literature suggests could be substantially larger than the commission savings alone.

E. Component 5: Lower Shipping and Logistics Costs

Original Ratio: 0.19% of vehicle price.[85]

Mechanism: Direct-sales models enable optimized hub-to-consumer delivery routes, reducing reliance on the multi-step manufacturer-to-dealer-to-consumer logistics chain that characterizes the franchised-dealer system. Modern destination fees have also increased substantially, with some domestic trucks carrying destination charges as high as $2,295.[86]

Calculation:

S₅ = ATP × 0.0019

S₅ = $50,000 × 0.0019 = $95

1. Assessment

This component is the smallest in the Goldman Sachs framework and likely the most conservative. Current destination fees—generally ranging from $995 to $2,295—substantially exceed the $95 savings estimated here. That gap suggests the original Goldman Sachs ratio may understate potential logistics savings. Destination fees, however, include transportation costs from the assembly plant to the retail network that would exist under any distribution model. The marginal savings from logistics optimization therefore remain modest relative to the other components.

At the same time, several modern logistics factors suggest that the $95 estimate likely understates the full inefficiency associated with dealer-based routing.

First, destination fees have risen sharply—from an average of about $839 in 2011 to a current range of roughly $995–$2,295, with a midpoint near $1,200. These fees are typically equalized across buyers of a given model regardless of distance from the assembly plant. That structure effectively cross-subsidizes transportation costs and reflects the logistical constraints of routing vehicles through a dispersed network of roughly 17,000 dealer lots, rather than through a smaller number of regional delivery hubs. A hub-based distribution system could price logistics more directly based on distance. If even 15%–25% of the average destination fee reflects this structural inefficiency, the avoidable cost would fall in the range of roughly $180–$300 per vehicle.^[87]

Second, dealer-to-dealer trades—in which a vehicle is transported from one dealership to another to match a buyer’s desired configuration—represent pure logistics waste that would not exist under a build-to-order or centralized-inventory model. Transfer fees for these trades range from $0 for short intra-market swaps to several hundred dollars for cross-state transports, with documented examples around $195 per transaction. Although no authoritative source quantifies the share of transactions involving dealer trades, the practice is common enough to support a dedicated logistics infrastructure.^[88]

Third, direct-to-consumer operators demonstrate that alternative logistics systems can operate at scale. Carvana, for example, runs a hub-and-spoke distribution network with vertically integrated transportation, reconditioning, and customer delivery. The company reports per-unit cost advantages of roughly $1,200–$2,600 relative to traditional dealer-intermediated distribution.^[89]

Taken together, these factors suggest that a reasonable but nonconservative estimate of logistics savings would fall in the range of $200–$400 per vehicle (roughly 0.40%–0.80% of ATP), compared with the $95 estimate produced by ratio scaling. This appendix nonetheless retains the $95 figure as the baseline, because the original Goldman Sachs derivation is not publicly available in sufficient detail to determine which logistics costs it already incorporated. The $200–$400 range should therefore be understood as a plausible upper bound reflecting destination-fee inefficiencies, dealer-trade logistics waste, and the demonstrated efficiency of hub-based direct-delivery models.

2. Cross-check

An additional cross-check for the savings estimate comes from Tesla’s operating experience as a direct-sales manufacturer. Bloomberg Intelligence reports that Tesla’s SG&A expense per delivery was $2,651 per vehicle in 2023, about 72% higher than Stellantis.[90] This figure reflects the cost of maintaining company-owned showrooms, mobile-service crews, and vertically integrated delivery logistics.

These figures provide a practical benchmark against the theoretical savings estimate, demonstrated in Table III. The resulting net efficiency gain of roughly $2,036 per vehicle indicates that even after accounting for the substantial costs of operating a vertically integrated retail and service network, direct sales can remain more cost-efficient than the franchised-dealer model.

APPENDIX TABLE III: Tesla Benchmark Against Theoretical Savings

[91]

The $2,651 figure also reflects Tesla’s continued buildout of its retail, service, and delivery infrastructure. As the network matures, economies of scale could reduce SG&A costs on a per-vehicle basis.

For policy purposes, the key question is not whether direct sales always produce lower costs in every circumstance. The relevant question is whether state law should categorically prohibit manufacturers from attempting a distribution model that economic evidence suggests can produce measurable efficiencies. Even the $2,036 net estimate represents a meaningful cost reduction for consumers.

V. Methodological Limitations and Caveats

The reconstruction presented in this appendix provides a conservative estimate of the potential distribution-cost savings from direct-sales models, but several methodological limitations should guide how the results are interpreted. The analysis relies on ratio scaling from the original Goldman Sachs framework, rather than a fully bottom-up re-derivation using current dealer cost data. It also evaluates the distribution channel in partial equilibrium, holding broader market conditions constant. In addition, structural changes in the automotive industry over the past quarter-century—including dealer consolidation, digital retail tools, and changes in product mix—may affect how the original ratios translate to current market conditions. The subsections below discuss these limitations and explain why, despite them, the estimates presented here likely understate rather than overstate the cost of mandatory dealer intermediation.

A. Linear Scaling of Cost Ratios

The methodology applies the original percentage ratios from the 2000 Goldman Sachs analysis to modern vehicle prices, implicitly assuming that the cost components scale proportionally with the average transaction price. In practice, some components may scale nonlinearly. Logistics costs, for example, may include larger fixed components, while savings from improved supply-demand matching could increase as vehicles become more complex. The directional bias of this assumption likely produces understatement. Over the past 25 years, interest rates, commercial real-estate costs, and vehicle complexity have all increased, factors that would tend to raise the cost of maintaining a large dealership network. A fully bottom-up reconstruction using current cost structures would therefore likely produce estimates at least as large as the ratio-scaled figures presented here.

B. Industry Changes Since 2000

The automotive industry has changed significantly since the Goldman Sachs report. Dealer consolidation may have produced some efficiency gains within the franchise system. Digital shopping tools have reduced some consumer information frictions. At the same time, the shift in product mix toward larger vehicles, particularly SUVs and trucks, has altered cost structures across manufacturing and distribution. These structural changes could influence the magnitude of distribution-cost savings in either direction.

C. Original Ratios Not Re-Derived

The 2026 estimates apply the original Goldman Sachs percentage decomposition to updated price data, rather than independently re-estimating each component using current cost structures. Future research should attempt a full bottom-up reconstruction using contemporary dealer financial data, such as the National Automobile Dealers Association’s “Annual Financial Profile of America’s Franchised New-Car Dealerships,” rather than relying solely on ratio extrapolation.

D. Partial-Equilibrium Framework

The estimates describe the savings associated with transitioning a single manufacturer’s distribution model while holding other market conditions constant. A system-wide transition to direct distribution could generate general-equilibrium effects, including changes in competitive dynamics, entry and exit patterns, and manufacturer pricing strategies. These second-order effects could be substantial and are ambiguous in direction.

E. Inflation Adjustment

The reconstruction applies the original Goldman Sachs percentage ratios to the 2026 average transaction price of $50,000 rather than simply inflating the original dollar figures from 2000. This approach is appropriate because cumulative consumer price index inflation from 2000 to 2025 is roughly 92%.[92] The increase in the average transaction price from $26,000 to about $50,000 therefore largely reflects general inflation, with only modest real appreciation. Applying the original ratios to the modern transaction price produces estimates broadly equivalent to an inflation-adjusted calculation.

VI. Conclusion

The reconstruction presented in this appendix is deliberately conservative at every step. It applies simple ratio scaling from the original Goldman Sachs framework, adopts the lowest plausible inputs where ranges exist, and explicitly flags areas of uncertainty. Under these cautious assumptions, the floor estimate is $3,934 per vehicle (7.87% of ATP), derived by taking the lowest value in each component range. Because cumulative CPI inflation from 2000 to 2025 is roughly 92%, the increase in the average transaction price from $26,000 to $50,000 largely reflects general price-level changes. Even the conservative floor therefore implies that the real cost of mandatory dealer intermediation is at least as large today as when Goldman Sachs first estimated it in 2000 and the DOJ documented it in 2009.

The conservative floor, however, is unlikely to represent the most realistic estimate. The bottom-up evidence consistently points in one direction: current market conditions have made the physical dealer infrastructure more expensive to maintain in real terms, not less. Interest rates are higher than in 2000, commercial real estate costs have risen, and vehicle complexity has increased. Where components could be reconstructed using contemporary data, the bottom-up figures generally met or exceeded the ratio-scaled estimates—inventory carrying costs of $1,045–$1,105, facility overhead of $745–$1,152, commissions of $449–$735, and shipping and logistics savings of up to $400 per vehicle. Using the upper end of these component ranges yields an estimate of roughly $4,992 per vehicle (9.87% of ATP). Even that figure likely understates the full economic burden of mandatory dealer intermediation because it excludes welfare costs—such as bargaining frictions, consumer time costs, and distributional harms—that the academic literature identifies as economically significant but that fall outside the Goldman Sachs accounting framework.

[1] See Daniel A. Crane, Tesla, Dealer Franchise Laws, and the Politics of Crony Capitalism, 101 Iowa L. Rev. 573 (2016),  https://repository.law.umich.edu/articles/1721.

[2] These figures represent the per-vehicle cost based on an average purchase price of $50,000, as discussed infra.

[3] See Gerald R. Bodisch, Economic Effects of State Bans on Direct Manufacturer Sales to Car Buyers, U.S. Dep’t Just. (2009), https://www.justice.gov/atr/economic-effects-state-bans-direct-manufacturer-sales-car-buyers.

[4] See Cox Auto. Inc., Kelley Blue Book Report: New-Vehicle Average Transaction Price Hits Record High in September, Surges Past $50,000 for the First Time Ever (Oct. 13, 2025), https://www.coxautoinc.com/insights-hub/sept-2025-atp-report.

[5] See Jonathan Gregory, Wage Gains Boost New-Vehicle Affordability in November Despite Higher Rates, Cox Auto. Inc. (Dec. 15, 2025), https://www.coxautoinc.com/insights-hub/nov-2025-vai (noting that purchasing a new vehicle requires about 36.3 weeks of income—still historically high, despite recent affordability improvements following pandemic-driven price increases).

[6] See Roger D. Blair et al., Brief of Legal and Economic Scholars to the Georgia Supreme Court in Lucid v. Georgia, Int’l Ctr. L. & Econ. (July 10, 2025), https://laweconcenter.org/resources/brief-of-legal-and-economic-scholars-to-the-georgia-supreme-court-in-lucid-v-georgia.

[7] Id.

[8] James M. Rubenstein, Making and Selling Cars: Innovation and Change in the U.S. Automotive Industry 188 (Johns Hopkins Univ. Press 2008).

[9] See Patrick Manzi, 2024 Annual Financial Profile of America’s Franchised New-Car Dealerships, Nat’l Auto. Dealers Ass’n (2024), https://www.nada.org/media/4695/download.

[10] Julie Walker, How Many U.S. Dealership Owners Will We Have in 2050?, Kerrigan Advisors (Mar. 30, 2025), https://www.kerriganadvisors.com/in-the-news/how-many-u-s-dealership-owners-will-we-have-in-2050.

[11] Mass. State Auto. Dealers Ass’n, Inc. v. Tesla Motors MA, Inc., 15 N.E.3d 1152 (Mass. 2014), https://law.justia.com/cases/massachusetts/supreme-court/2014/sjc-11545.html.

[12] See George J. Stigler, The Theory of Economic Regulation, 2 Bell J. Econ. & Mgmt. Sci. 3 (1971).

[13] Bodisch, supra note 3.

[14] 2024 Car Buyer Journey Study, Cox Auto. Inc. (2025), https://www.coxautoinc.com/wp-content/uploads/2025/01/2024-Car-Buyer-Journey-Study-Research-Summary.pdf.

[15] Id.

[16] Id. at 4.

[17] Id.

[18] As discussed below, this analysis yields estimated savings ranging from a conservative baseline of 7.87% per vehicle to 9.87% per vehicle. The analysis assumes a $50,000 new-vehicle price. Table II illustrates how these estimates translate across a range of vehicle price points.

[19] Id.

[20] Id. at 5-6.

[21] Cox Auto Inc., supra note 4 (noting that recent increases in average vehicle prices largely track overall inflation; see also Appendix A, § 7.E); Nora Eckert, Automakers Have Resisted Raising Car Prices Because of Tariffs. That Might Not Last, Reuters (Sept. 18, 2025), https://www.reuters.com/business/media-telecom/automakers-have-resisted-raising-car-prices-because-tariffs-that-might-not-last-2025-09-18 (reporting that 2025 U.S. tariffs—including 25% duties on imported vehicles and auto parts—have increased costs for automakers and suppliers, potentially adding about $2,300 per vehicle); The Latest Car Tariff Information, Kelley Blue Book (Feb. 25, 2026), https://www.kbb.com/tariffs/#:~:text=New%20cars%3A%20Shoppers%20can%20expect,cars%20less%20affordable%20for%20consumers (estimating that tariffs could raise prices by as much as $6,000 on vehicles priced under $40,000 as duties phase in); Lawrence Ulrich, Trump’s Auto Tariffs Are Turning New-Car Shopping into a Wild Ride, Car & Driver (Aug. 5, 2025), https://www.caranddriver.com/news/a65592759/trump-auto-tariffs-new-car-shopping-outlook.

[22] See Thomas L. Hidder, Floor-Plan Financing for Auto Dealers: Trends, Structures & What’s Changing, Harney Partners (Nov. 24, 2025), https://harneypartners.com/floor-plan-financing-for-auto-dealers; Prime Rate Information, Bank Am. (Dec. 11, 2025), https://newsroom.bankofamerica.com/content/newsroom/home/prime-rate-information.html (reporting a prime rate of 6.75% effective December 2025; this brief uses an approximate 6% rate in the bottom-up reconstruction as a conservative assumption; see Appendix A).

[23] See Cox Auto. Inc., New-Vehicle Inventory Returns to Pre-Tariff Levels as EV Sales Accelerate and Incentives Rise (Aug. 14, 2025), https://www.coxautoinc.com/insights-hub/july-2025-new-vehicle-inventory.

[24] See Appendix A § 5.

[25] See Optimum Info, Dealer Financial Analysis Report—Q1 2025, https://optimuminfo.com/resources/blogs/dealer-financial-analysis-report-q1-2025 (last visited Feb. 27, 2025).

[26] Pike v. Bruce Church, Inc., 397 U.S. 137, 142 (1970).

[27] See Fed. Trade Comm’n, FTC Staff: Missouri and New Jersey Should Repeal Their Prohibitions on Direct-to-Consumer Auto Sales by Manufacturers (May 16, 2014), https://www.ftc.gov/news-events/news/press-releases/2014/05/ftc-staff-missouri-new-jersey-should-repeal-their-prohibitions-direct-consumer-auto-sales.

[28] Id.

[29] See Healy v. Beer Inst., 491 U.S. 324, 336 (1989) (“The Commerce Clause … precludes the application of a state statute to commerce that takes place wholly outside of the State’s borders, whether or not the commerce has effects within the State.”); Crane, supra note 1 (arguing that state dealer-franchise laws effectively dictate the distribution structures out-of-state manufacturers must adopt to access a state’s market); Brannon P. Denning, Extraterritoriality and the Dormant Commerce Clause: A Doctrinal Post-Mortem, 73 La. L. Rev. 979 (2013).

[30] Mass. State Auto. Dealers Ass’n, Inc. v. Tesla Motors MA, Inc., 469 Mass. 675, 15 N.E.3d 1152 (2014).

[31] See Ill. Auto Dealers Ass’n v. Office of the Ill. Sec’y of State, No. 1-23-0100 (Ill. App. Ct. Aug. 23, 2024); Sarah J. Reusche, Illinois Appellate Court Affirms Direct-to-Consumer Auto Sales, Lavelle L. (Sept. 9, 2024), https://www.lavellelaw.com/illinois-appellate-court-affirms-direct-to-consumer-auto-sales.

[32] Blair et al., supra note 6.

[33] See John Smart, Rivian’s $4.6 Million Ballot Gambit to Crack Washington’s Tesla Sales Monopoly, WebProNews (Jan. 27, 2026), https://www.webpronews.com/rivians-4-6-million-ballot-gambit-to-crack-washingtons-tesla-sales-monopoly; Lisa Stiffler , The ‘Tesla exemption’ No More: Rivian and Lucid Break Through Washington State’s Dealership Wall, GeekWire (Mar. 12, 2026), https://www.geekwire.com/2026/the-tesla-exemption-no-more-rivian-and-lucid-break-through-washington-states-dealership-wall/.

[34] See Andrew J. Hawkins, Rivian Calls Ohio’s Ban on Direct Car Sales ‘Irrational in the Extreme’ in New Lawsuit, The Verge (Aug. 4, 2025), https://www.theverge.com/news/718186/rivian-ohio-lawsuit-direct-sales-ban.

[35] See Matt Wilson, Rivian Pushes for Direct Sales as Iowa Weighs Changes to Dealership Laws, Drive Tesla (Feb. 13, 2026), https://driveteslacanada.ca/news/rivian-pushes-for-direct-sales-as-iowa-weighs-changes-to-dealership-laws (discussing Iowa Senate Study Bill 3067 (proposed 2025)); Carrigan Woodson, South Carolina’s Direct-to-Consumer Car Sales Bill Ran Out of Time in 2025. Should It Be Revived in 2026?, Palmetto Promise (Aug. 5, 2025), https://palmettopromise.org/south-carolinas-direct-to-consumer-car-sales-bill-ran-out-of-time-in-2025-should-it-be-revived-in-2026.

[36] See Crane, supra note 1.

[37] See 15 U.S.C. § 1221.

[38] See Nat’l Auto. Dealers Ass’n v. FTC, No. 24-60013 (5th Cir. Jan. 27, 2025) (vacating 16 C.F.R. pt. 463 for failure to comply with advance-notice requirements under § 18 of the Federal Trade Commission Act).

[39] Thomas Randolph Beard & George S. Ford, State Automobile Franchise Laws: Public or Private Interests?, SSRN (Sept. 12, 2016).

[40] Gary Lapidus, eAutomotive: Gentleman Start Your Search Engines, Goldman Sachs (2000). The original Lapidus report is unavailable. This analysis therefore relies on Bodisch, supra note 3, which cites Lapidus and Charles H. Fine & Daniel M. G. Graff, Automotive Industry: Internet-Driven Innovation and Economic Performance, in The Economic Payoff from the Internet Revolution (Robert E. Litan & Alice M. Rivlin eds., 2001), https://archive.org/details/economicpayofffr0000unse.

[41] Bodisch, supra note 3.

[42] In 2000, the average corporate-bond yield was 7.6%; by February 2026, it had fallen to 5.3%. See Moody’s Seasoned Aaa Corporate Bond Yield (AAA), FRED, Fed. Rsrv. Bank St. Louis, https://fred.stlouisfed.org/series/AAA (last visited Feb. 25, 2026).

[43] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[44] Cox Auto Inc., supra note 4.

[45] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[46] Bodisch, supra note 3.

[47] Id.

[48] Cox Auto Inc., supra note 4.

[49] The 2000 Goldman Sachs report expressed savings as percentage shares of the average transaction price, rather than reporting individual cost inputs. The variables marked “N/A”—including floorplan rates, incentive levels, and dealership operating costs—were implicit in the original analysis but were not reported as standalone figures. Current values are provided here to show that changes since 2000 support, rather than undermine, the original ratios.

[50] See infra notes 28-34 and accompanying text.

[51] Cox Auto. Inc., New-Vehicle Inventory Returns to Pre-Tariff Levels as EV Sales Accelerate and Incentives Rise (July 2025), https://www.coxautoinc.com/insights-hub/july-2025-new-vehicle-inventory.

[52] Cox Auto Inc., supra note 4.

[53] $708 vs. $250: What Dealers Really Spend Per Car Sold, Big Time Advert. & Mktg. (2025), https://www.gowithbigtime.com/blog/C/2025/10/01/708-vs-250-What-Dealers-Really-Spend-Per-Car-Sold-177.

[54] The per-vehicle commission derives from Presidio-NCM benchmark data reporting average new-vehicle gross profit of $2,247 for full-year 2024, applied to an industry-standard 20–25% front-end commission rate ($2,247 × 0.20–0.25 = $449–$562 per vehicle). See James Hickey, Dealer Profitability Improved at End of 2024, Digital Dealer (Feb. 28, 2024), https://digitaldealer.com/news/auto-dealership-profitability-improved-at-end-of-2024/163923 (reporting Presidio-NCM benchmark data); AutoFinder, How Much Commission Does a Car Salesperson Make?, https://autofinder.com/insights/how-much-commission-does-a-car-salesman-make (reporting commission rates of 20–30% of gross profit); Mark McDonald, Car Salesman Confidential: How to Get Paid, MotorTrend (Feb. 14, 2014), https://www.motortrend.com/features/car-salesman-confidential-how-to-get-paid (reporting typical commissions of about 25% of gross profit plus backend compensation); Lawrence Hodge, How Much Car Salespeople Make in Commissions, Jalopnik (Feb. 20, 2025), https://www.jalopnik.com/1795349/how-much-car-salespeople-make-commissions (reporting commissions of roughly 20% on new-car sales).

[55] Renee Valdes, Destination Charges and Dealer Fees Explained, Autotrader (Apr. 30, 2024), https://www.autotrader.com/car-shopping/new-car-delivery-or-destination-charges-explained.

[56] Nat’l Auto. Dealers Ass’n, NADA Data: Annual Financial Profile of America’s Franchised New-Car Dealerships (2024), https://www.nada.org/nada/nada-data (reporting average dealership sales of $73.3 million with net margins of 1–2%, implying non-vehicle-acquisition operating expenses of roughly $500,000 per month or more).

[57] Cox Auto Inc., supra note 4.

[58] Id.

[59] Nathan Paulus, Average New Car Price in 2026: $48,841, MoneyGeek (Dec. 29, 2025), https://www.moneygeek.com/resources/average-price-of-a-new-car.

[60] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[61] Bodisch, supra note 3.

[62] Cox Auto Inc., supra note 4.

[63] Charles Murry & Yiyi Zhou, Consumer Search and Automobile Dealer Colocation, 66 Mgmt. Sci. 1909 (2020), https://doi.org/10.1287/mnsc.2019.3307.

[64] Florian Zettelmeyer, Fiona Scott Morton & Jorge Silva-Risso, How the Internet Lowers Prices: Evidence from Matched Survey and Automobile Transaction Data, 43 J. Mktg. Rsch. 168 (2006), https://doi.org/10.1509/jmkr.43.2.168.

[65] Cox Auto Inc., supra note 4.

[66] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[67] Bodisch, supra note 3.

[68] See Optimum Info, Dealer Financial Analysis Report—Q1 2025 (2025), https://optimuminfo.com/resources/blogs/dealer-financial-analysis-report-q1-2025; see also Harney Partners, Floor-Plan Financing for Auto Dealers: Trends, Structures & What’s Changing (2025), https://harneypartners.com/floor-plan-financing-for-auto-dealers; Bank Am., Prime Rate Information (reporting a prime rate of 6.75% effective December 2025), https://newsroom.bankofamerica.com/content/newsroom/home/prime-rate-information.html.

[69] Cox Auto. Inc., New-Vehicle Inventory Returns to Pre-Tariff Levels as EV Sales Accelerate and Incentives Rise (July 2025), https://www.coxautoinc.com/insights-hub/july-2025-new-vehicle-inventory.

[70] The $15-per-day non-interest holding cost (Cₒₚ) is an estimate. No published source reports a per-vehicle, per-day figure for noninterest inventory-carrying costs. This estimate derives from aggregate dealership operating data. National Automobile Dealers Association financial profiles report average dealership payroll of about $5.39 million annually—roughly $449,000 per month—with payroll typically accounting for about 40%–50% of operating expenses. That implies total operating costs of roughly $900,000–$1.1 million per month per dealership. See Nat’l Auto. Dealers Ass’n, NADA Data: Annual Financial Profile of America’s Franchised New-Car Dealerships (2024), https://www.nada.org/nada/research-data/nada-data. To remain conservative, this brief assumes total operating costs of about $500,000 per month, only modestly above payroll and therefore likely understating the full overhead associated with maintaining dealership facilities and inventory. A typical dealership lot holds about 150–250 vehicles. See Andy Kalmowitz, Nearly 3 Million Cars Are Sitting on Dealer Lots, Jalopnik (July 12, 2024), https://www.jalopnik.com/nearly-3-million-cars-are-sitting-on-dealer-lots-1851587949 (reporting 2.89 million vehicles across roughly 17,000 lots). Allocating a conservative share of dealership operating costs to inventory maintenance—including insurance, lot upkeep, security, and depreciation exposure—yields an estimated holding cost of roughly $10–$20 per vehicle per day, assuming that about 10%–20% of dealership operating costs relate to maintaining inventory on dealership lots. This brief uses the midpoint of that range.

[71] Bd. of Governors Fed. Rsrv. Sys., Bank Prime Loan Rate [DPRIME], FRED, Fed. Rsrv. Bank St. Louis, https://fred.stlouisfed.org/series/DPRIME (last visited Feb. 25, 2026).

[72] Gérard P. Cachon & Marcelo Olivares, Drivers of Finished-Goods Inventory in the U.S. Automobile Industry, 56 Mgmt. Sci. 202 (2009), https://doi.org/10.1287/mnsc.1090.1095; Marcelo Olivares & Gérard P. Cachon, Competing Retailers and Inventory: An Empirical Investigation of General Motors’ Dealerships in Isolated U.S. Markets, 55 Mgmt. Sci. 1586 (2009), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1946447.

[73] Adam Copeland, Wendy E. Dunn & George Hall, Inventories and the Automobile Market, 42 RAND J. Econ. 121 (2011), https://www.jstor.org/stable/23046792.

[74] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[75] Nat’l Auto. Dealers Ass’n, NADA Data: Annual Financial Profile of America’s Franchised New-Car Dealerships (2024), https://www.nada.org/nada/research-data/nada-data.

[76] Nat’l Auto. Dealers Ass’n, NADA Market Beat (2025), https://www.nada.org/nada/research-data/market-beat (reporting a seasonally adjusted annual rate of approximately 16.2–16.3 million units).

[77] Mark C. Anderson, Rajiv D. Banker & Surya Janakiraman, Are Selling, General, and Administrative Costs “Sticky”?, 41 J. Acct. Rsch. 47 (2003), https://www.jstor.org/stable/3542244.

[78] Francine Lafontaine & Fiona Scott Morton, Markets: State Franchise Laws, Dealer Terminations, and the Auto Crisis, 24 J. Econ. Persp. 233 (2010), https://doi.org/10.1257/jep.24.3.233.

[79] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[80] See Cox Auto. Inc., supra note 21.

[81] Id.

[82] Ambarish Chandra, Sumeet Gulati & James Sallee, Who Loses When Prices Are Negotiated? An Analysis of the New Car Market, 65 J. Indus. Econ. 235 (2017), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2602526.

[83] Meghan R. Busse, Jorge Silva-Risso & Florian Zettelmeyer, $1,000 Cash Back: The Pass-Through of Auto Manufacturer Promotions, 96 Am. Econ. Rev. 1253 (2006), https://www.nber.org/system/files/working_papers/w10887/w10887.pdf.

[84] Andreas Grunewald et al., Auto Dealer Loan Intermediation: Consumer Behavior and Competitive Effects, Nat’l Bureau Econ. Rsch., Working Paper No. 28136 (2020), https://doi.org/10.3386/w28136.

[85] Lapidus, supra note 40, as cited in Bodisch, supra note 3.

[86] Renee Valdes, Destination Charges and Dealer Fees Explained, Autotrader (Apr. 30, 2024), https://www.autotrader.com/car-shopping/new-car-delivery-or-destination-charges-explained.

[87] Mike Monticello, The Truth About Destination Fees, Consumer Reps. (Aug. 15, 2023), https://www.consumerreports.org/cars/buying-a-car/the-truth-about-destination-fees-a1615480982; Chris Hardesty, What Are Destination Charges?, Kelley Blue Book (May 2, 2025), https://www.kbb.com/car-advice/what-are-destination-charges.

[88] See, e.g., Powers Swain Chevrolet, Will a Car Dealership Drop the Transfer Fee from Another Dealership? (July 5, 2024), https://www.pschevy.com/blog/2024/july/5/will-a-car-dealership-drop-the-transfer-fee-from-another-dealership.htm (reporting a $194.50 transfer fee for a dealer-to-dealer vehicle locate).

[89] See Carvana Co., Cost Structure Details (Nov. 2023), https://investors.carvana.com/~/media/Files/C/Carvana-IR/documents/cost-structure-details.pdf (reporting per-unit cost advantages of $1,200–$2,600 from vertically integrated inbound transportation, reconditioning, and customer delivery); see also infra § 6 (Cross-Check) (comparing Tesla’s SG&A per delivery with legacy OEMs).

[90] Bloomberg Intelligence, Retail Auto Dealers Deep Dive (June 5, 2024), https://www.madaonline.com/sites/default/files/Bloomberg_Intelligence_Retail_Auto_Dealers_June_2024.pdf.

[91] Id.

[92] See U.S. Bureau Lab. Stat., CPI Inflation Calculator, https://www.bls.gov/data/inflation_calculator.htm (last visited Feb. 25, 2026) (reporting that $26,000 in 2000 equals $50,098.06 in 2026, a cumulative increase of more than 92%).

Executive Summary

Policymakers have proposed capping credit card interest rates to provide relief to consumers. A recent Vanderbilt Policy Accelerator study argues that banks’ margins are sufficient to absorb such caps while continuing to lend profitably, yielding net consumer savings.

That conclusion relies on a static framework that asks only whether lending remains arithmetically viable under a cap. It does not consider how lenders, borrowers, and competing providers will respond when risk-based pricing is constrained. Economic theory and evidence show that credit markets adjust along other margins: supply contracts for higher-risk borrowers, shifts toward lower-risk segments, and migrates to substitute or informal channels.

The Vanderbilt study also rests on unsupported assumptions. It assumes issuers will uniformly price at the cap, even though that could raise costs for prime borrowers and turn a ceiling into a de facto floor. It assumes banks will offset losses by cutting advertising, despite likely effects on competition and innovation. And it evaluates viability using accounting returns, rather than the risk-adjusted returns that drive capital allocation.

A proper welfare analysis must account for borrowers pushed out of the market, substitution into costlier alternatives, fee shifting, and reduced innovation. Evidence from prior reforms and international experience confirms these effects.

Credit markets are dynamic. Price ceilings do not eliminate risk or create credit; they reallocate it. The relevant question is whether restricting risk-based pricing improves welfare once these adjustments are considered. The evidence suggests it does not.

I. Introduction

Policymakers across the political spectrum have recently proposed bills to impose price controls on credit card interest rates. In February 2025, Sens. Josh Hawley (R-Mo.) and Bernie Sanders (I-Vt.) introduced S.381, the 10 Percent Credit Card Interest Rate Cap Act.[1] Sanders argued that the legislation would provide “desperately needed financial relief” to working families struggling to pay their bills.[2]

Economists have long viewed such proposals with skepticism. Price controls are blunt instruments, and their unintended but predictable effects often outweigh any intended benefits. It is therefore notable that in September 2025 the Vanderbilt Policy Accelerator published a study concluding that interest-rate caps would benefit consumers.[3]

This brief evaluates that study’s methodology, analysis, and conclusions. The study asks whether banks’ current profit margins are sufficient to absorb a mandated rate cut—and answers yes. But it does not address the more relevant question: how banks, borrowers, and substitute lenders will respond to a binding price ceiling. That question has been studied extensively, and the evidence does not favor rate-cap proponents.

II. The Limits of Static Analysis

The Vanderbilt study, authored by Brian Shearer—a lawyer and former assistant director at the Consumer Financial Protection Bureau (CFPB)—uses comparative static analysis to estimate the effects of capping credit card interest rates at 10%, 15%, or 18%. It relies on data from a Federal Reserve Bank of New York report by Itamar Drechsler (hereinafter the “Drechsler Report”), which decomposes credit card profitability into key components: interest spread, charge-off losses, interchange revenue, rewards costs, fee revenue, and operating expenses. The data cover borrower and transactor accounts across 14 FICO score bins, ranging from 600 to 850.[4]

Shearer replaces the observed interest spread with a hypothetical spread consistent with each proposed cap—for example, a 5% spread under a 10% cap, assuming a 5% federal funds rate. He then recalculates return on assets (ROA) and asks whether it remains positive.

The results appear striking. Under a 15% cap, nearly every FICO tier generates positive returns without adjustments to rewards or operating costs, suggesting that existing margins could absorb a meaningful rate reduction. Under a 10% cap, ROA turns negative in tiers below FICO 780 before adjustments. Shearer argues, however, that banks could restore profitability by cutting rewards in those tiers and trimming what he characterizes as “bloated advertising budgets,”[5] while still leaving consumers better off because interest savings would exceed rewards reductions by at least three to one.

These calculations are internally consistent but rest on a critical assumption: that a federal price control on credit card rates would produce no behavioral response. The analysis assumes no credit rationing, no product substitution, no fee shifting, and no portfolio rebalancing. That assumption is not merely contestable—it conflicts with extensive empirical evidence on how credit markets respond to binding price ceilings.

A sound policy analysis must also look beyond borrowers who remain in the market. Consumer welfare cannot be measured solely by comparing interest payments before and after a cap. It must account for borrowers who are rationed out of credit markets, the costs of shifting to substitute or informal credit, changes in fees and other less-salient pricing terms, reduced competition and innovation, and the reallocation of capital toward uncapped products. A policy that lowers rates for some borrowers while excluding others—or pushing them into inferior alternatives—may produce visible short-term savings while reducing overall welfare once these broader adjustments are considered. Because the Vanderbilt study focuses only on projected savings within the existing portfolio, it does not capture the full economic consequences of a binding cap.

III. Credit Rationing Under Price Controls

In a seminal 1981 paper, Andrew Weiss and Nobel Prize–winning economist Joseph Stiglitz showed that when lenders cannot adequately price risk—whether due to a binding usury cap or asymmetric information—markets do not clear through prices.[6] Instead, they clear through quantity rationing. When lenders cannot raise interest rates to reflect higher risk, they restrict the supply of credit to riskier borrowers.

Credit card markets follow this logic. A binding rate cap prevents lenders from charging risk-adjusted prices to higher-risk borrowers. Lenders respond by tightening nonprice terms: raising minimum credit score thresholds, lowering credit limits, shortening maturities, requiring additional collateral, or denying credit altogether. The result is both a reduction in total credit and a reallocation of credit toward lower-risk borrowers.

Shearer acknowledges this dynamic only at the margins. He notes, for example, that lending to borrowers with FICO scores of 600 and below may decline under a 10% cap because those accounts remain unprofitable even after eliminating rewards and cutting advertising. He treats this as a narrow and acceptable edge case, suggesting that “at that risk-level, we might see increases in other fees or a reduction in lending.”[7]

That framing understates the stakes. Even if the effects were limited to borrowers with FICO scores of 600 and below, excluding them from the credit card market would raise serious concerns. This group may represent “only 12% of Americans,” but dismissing granting their ability to access credit cards as resembling “predatory lending”[8] overlooks why their borrowing costs are higher in the first place: elevated default risk. Higher APRs reflect that risk; they do not create it. Nor does the possibility of debt cycles justify denying access to credit altogether. If anything, evidence on product substitution suggests that consumers denied mainstream credit often turn to more expensive or less regulated alternatives.

At bottom, this reasoning reflects a paternalistic premise: that individuals with low FICO scores are better off excluded from credit markets than allowed to borrow at risk-adjusted prices. That premise is both contestable and inconsistent with evidence on how consumers respond to credit constraints.

Rationing, moreover, is not confined to the lowest FICO tiers. A 2025 Federal Reserve Bank of New York study by Rajashri Chakrabarti et al. examined the introduction of 36% APR caps in several states and found that credit to higher-risk borrowers contracted, while credit to lower-risk borrowers expanded—exactly the reallocation predicted by theory.[9] The effects were concentrated among borrowers whose risk profiles implied market-clearing rates above the cap. If a 36% cap produces measurable rationing, a 10% or 15% federal cap would likely have far more pronounced effects.

A static accounting framework cannot capture these dynamics because it does not model behavior. Shearer’s approach shows only that bank profitability at each FICO tier is arithmetically consistent with lower interest rates. That is true in the same narrow sense that a restaurant’s gross margins might appear sufficient to absorb a price ceiling. What the arithmetic cannot show is whether the restaurant will remain open, change its menu, shift costs, or exit the market. Credit markets are not vending machines that dispense loans at regulator-set prices.

IV. Substitution Effects and Credit Migration

The Vanderbilt study largely ignores a central question: what happens to borrowers who lose access to credit cards—or face sharply reduced credit limits—under a binding rate cap? This is not a theoretical concern. A World Bank review of interest-rate caps across more than 70 countries identifies “migration to the least-constrained channel” as a pervasive and well-documented response.[10] When regulation targets a specific product, lenders adjust, and borrowers seek substitutes.

One such substitute is buy now, pay later (BNPL) financing—e.g., pay-in-four loans financed by merchants—as well as installment loans repaid over time with interest. The CFPB’s January 2025 report finds that about 21% of Americans with a credit score have used BNPL.[11] Use is heavily concentrated among lower-credit borrowers: “subprime” consumers (FICO scores of 580–619) account for 16% of BNPL loans, while “deep subprime” consumers (FICO scores of 300–579) account for 45%.[12] These are precisely the borrowers most likely to face credit restrictions under a binding cap on credit card interest rates.

BNPL, however, is an imperfect substitute for credit cards. Merchant offers of BNPL remain limited, and much more limited than the general prevalence of credit card acceptance. BNPL products also typically offer weaker fraud protections (for example, completely lacking zero-liability policies), along with more cumbersome dispute-resolution processes than credit card chargebacks. They also generally do not provide insurance benefits. In addition, pay-in-four products are usually limited to smaller purchases.

For larger purchases, borrowers may turn to installment loans. But annual percentage rates (APRs) on such loans for lower-FICO borrowers often match or exceed credit card rates. At the same time, many states have imposed caps on installment-loan APRs, which has reduced or eliminated their availability for the very consumers most likely to be rationed out of the credit card market.

Borrowers who cannot access credit cards, BNPL, or installment loans may turn to other sources, such as payday lending, which is often more expensive. If caps extend across all legal credit products—through a federal usury rule, for example—some borrowers will turn to illegal lending markets.

International experience underscores this risk. After Japan tightened rate caps under the 2006 revision of the Money Lending Business Act, consumer-finance balances fell sharply, while illegal lending expanded. Some illegal lenders reportedly charged rates as high as 10% per week—equivalent to an APR of 14,104%.[13] Korea’s experience is similar. After progressively lowering its statutory maximum rate to 20% by 2021, one study estimates that a single 2018 rate reduction pushed roughly 659,000 borrowers into informal or illegal credit channels.[14]

The Vanderbilt study’s failure to engage with these substitution effects is not a minor omission. It overlooks one of the most consistent empirical findings in the literature on price controls in credit markets.

V. Fee Shifting and Hidden Price Adjustments

When governments cap one source of revenue, lenders typically recover losses elsewhere. Fee shifting is a well-documented response to interest-rate restrictions. It can take many forms: higher annual fees, increased penalty and late fees, reduced benefits such as credit insurance, tighter grace periods, and higher minimum payment requirements that raise the effective cost of revolving balances.

Shearer argues that such fee shifting is unlikely, relying on a 2015 study by Sumit Agarwal, Souphala Chomsisengphet, Neale Mahoney, and Johannes Stroebel on the effects of the CARD Act. That study estimated that the act’s fee restrictions saved consumers $11.9 billion annually, with “no evidence of an offsetting increase in interest charges or a reduction in the volume of credit.”[15] But more recent research—using longer time horizons, richer data, and structural models unavailable to Agarwal et al.—reaches a different conclusion. The evidence now indicates that the CARD Act produced substantial fee shifting, altered competitive dynamics, restricted credit access for higher-risk borrowers, and imposed welfare costs that the study’s framework did not capture.

Agarwal et al. based their conclusions on a difference-in-differences model applied over a short observation window. Broader evidence tells a different story. The CFPB’s own 2013 report found that “[d]ue to the increase in both the incidence and average dollar amount of annual fees, consumers…paid an additional $475 million in annual fees in 2012.”[16]

Fee shifting has also intensified over time. Using CFPB data, the Consumer Bankers Association reports that total annual fees have doubled since 2015. While increased card issuance explains part of that growth, issuers also shifted the fee burden away from subprime borrowers and toward higher-credit tiers.[17] Lenders did not simply absorb lost revenue; they restructured pricing in ways that short-run models could not detect.

Agarwal et al. acknowledged these limits. With only two years of post–CARD Act data, they could not evaluate longer-run effects on market entry and exit, contract structure, or investment.[18] Subsequent evidence shows precisely those adjustments: rising annual fees, redesigned rewards programs, and changes in competitive behavior.

The most rigorous challenge to Agarwal et al. comes from Scott Nelson’s structural model of the credit card market.[19] Accounting for private information in repricing—the mechanism the CARD Act restricts—Nelson finds distributional effects absent from earlier work. Pricing became less responsive to both public and private measures of borrower risk, and price dispersion fell by roughly one-third, indicating more pooled pricing. Rates declined for higher-risk and less price-sensitive borrowers, while rising elsewhere—including for some borrowers with low FICO scores but favorable private risk profiles.

In other words, the CARD Act induced substantial fee shifting and cross-subsidization. It also contributed to partial market unraveling among relatively safe subprime borrowers, who were priced out of credit they would otherwise have received. Nelson further shows that part of the earlier study’s measured increase in “consumer surplus” reflects reduced lender profits—a transfer to remaining borrowers, not a net welfare gain. That transfer came at the expense of borrowers who exited the market entirely, whose losses are not captured in account-level data.[20]

Other studies corroborate these findings. Yiwei Dou, Julapa Jagtiani, Joshua Ronen, and Kenechukwu Maingi show that the ratio of credit limits to total available credit declined for subprime borrowers relative to higher-score groups after the CARD Act.[21] Crucially, this shift began with the legislation—not during the Great Recession—undermining claims that macroeconomic conditions drove the change. Song Han, Benjamin Keys, and Geng Li find similar effects in credit card solicitation data, documenting a contraction in supply to subprime borrowers.[22] Gregory Elliehausen and Simona Hannon find that restrictions on risk-management practices reduced credit card holding among higher-risk consumers and increased reliance on consumer finance loans in states where those loans remained viable.[23] In short, borrowers who lost access to credit cards shifted to costlier alternatives—a welfare loss not reflected in the original estimates.

Finally, Shearer contends that consumers choose credit cards based on salient features such as annual fees and rewards, limiting issuers’ ability to shift costs. But if consumers focus on visible terms, issuers have stronger incentives to adjust less salient ones: tighter grace periods, higher minimum payments, and more aggressive penalty structures. The study does not model these margins of adjustment.

VI. The Uniform Pricing Assumption

A further key assumption in the Vanderbilt study is that all issuers will charge the capped interest rate to all borrowers. Shearer describes this as a “conservative” assumption, reasoning that if banks charge less than the cap, consumers save even more. But the assumption is neither conservative nor empirically grounded.

It may hold for lower-FICO tiers, where current rates exceed any proposed cap. For higher-FICO borrowers—particularly those currently paying interest spreads in the 8–12% range—the assumption implies that rates would rise under a cap, not fall. Shearer acknowledges as much, noting that under an 18% cap, borrowers with FICO scores above 760 would pay roughly $7 billion more each year.

The assumption draws on the payday-lending literature, where lenders often charge the maximum permissible rate.[24] But that pattern reflects the structure of payday lending, which targets high-risk borrowers with few alternatives. In that context, the cap may approximate the minimum viable rate. Credit card markets are different. Issuers serve borrowers across a wide range of risk profiles and compete on multiple price and nonprice dimensions. By assuming uniform pricing at the cap, the study imports a model that does not fit the market it seeks to analyze.

If issuers continue to price below the cap for prime borrowers, the study overstates consumer savings. If they instead converge on the cap, then the policy functions as a price floor for low-risk borrowers—raising their borrowing costs while purporting to help consumers. Either way, the study’s savings estimates are unreliable.

VII. The Advertising Assumption

In FICO tiers where static ROA turns negative under a rate cap, the Vanderbilt study assumes that banks will restore profitability by cutting advertising expenditures. That assumption is flawed in several respects.

First, the study assumes that marketing accounts for 20% of expenses across all issuers and FICO tiers.[25] This figure derives from a single observation about Capital One’s expense structure in the Drechsler Report—and refers to “marketing,” not advertising.[26] That is a thin empirical basis for a system-wide assumption with significant analytical consequences.

Second, marketing is not a discretionary expense that banks can reduce without tradeoffs. It is an endogenous investment in customer acquisition and retention. Marketing allows issuers to inform consumers about existing products and introduce new ones. A uniform reduction in marketing might leave relative market shares unchanged, but only by dampening competition and slowing product differentiation. That outcome appears implicit in the study’s assumption.

Reduced marketing would also affect innovation. If expected returns fall, banks will scale back investment in new products and features. Lower marketing spend reduces product adoption and revenue, weakening incentives to invest in improvements such as fraud protection, insurance benefits, and digital account management tools. These are not marginal features; they are central to the value proposition of modern credit cards.

More fundamentally, the claim that advertising budgets are “bloated” lacks a clear benchmark. The assumption that banks will absorb the effects of rate caps primarily through advertising cuts reflects the same static logic that underpins the broader analysis. In reality, banks would adjust along multiple margins—credit rationing, product substitution, and fee restructuring among them. Assigning the entire adjustment burden to marketing is inconsistent with both theory and evidence.

VIII. Capital Allocation and Opportunity Cost

A final—and more fundamental—weakness in the Vanderbilt study lies in its treatment of capital allocation and opportunity cost. The study assumes that a positive ROA is sufficient to sustain lending within a given FICO tier. If projected ROA remains above zero—or above the federal funds rate—lending is deemed “viable.” That framing misstates how banks allocate capital and leads to an overly optimistic view of credit card lending under rate caps.

Banks do not evaluate business lines against a zero-profit threshold. They allocate scarce regulatory capital based on risk-adjusted return on capital (RAROC).[27] The relevant question is not whether a product generates a positive accounting return, but whether it generates a return commensurate with the capital it consumes, the volatility of its cash flows, its correlation with macroeconomic risk, and the returns available from alternative uses of capital.

Credit cards score poorly on these dimensions. They are unsecured, operationally intensive, and highly procyclical. Charge-off rates rise sharply in downturns, reflecting both income volatility and the absence of collateral.[28] Under Basel standards and U.S. implementing regulations, unsecured revolving credit carries materially higher capital charges than many secured products.[29] Banks must hold capital against unexpected losses, not just expected losses, precisely because these portfolios exhibit elevated loss volatility.[30]

Rate caps compress the risk premium available to compensate for that volatility. If a 15% cap reduces projected ROA in binding tiers to roughly 2%, as the Vanderbilt study suggests, that figure must be evaluated in context. A 2% ROA on a high-risk, high-volatility unsecured product is not equivalent to a 2% ROA on secured lending or fee-based activities. Nor is it equivalent to alternative uses of capital once returns are evaluated on a risk-adjusted basis.

The federal funds rate is also the wrong benchmark. Banks cannot freely redeploy capital into reserves and earn the policy rate without regard to capital requirements, liquidity rules, leverage constraints, and shareholder expectations.[31] The relevant comparison is the marginal risk-adjusted return available elsewhere. If capped returns fall below a bank’s internal hurdle rate—typically set above its cost of equity and calibrated to portfolio risk—capital will have to be reallocated.[32]

That reallocation need not be dramatic to matter. Even modest shifts toward prime borrowers, secured lending, commercial credit, or fee-based services would reduce supply in the segments where caps bind most tightly. The Vanderbilt framework assumes capital remains in place so long as returns are positive. In practice, capital flows to its highest risk-adjusted use.[33]

Static ROA analysis also misses the dynamic nature of credit card lending. These portfolios depend on sustained investment in underwriting, fraud detection, rewards systems, customer acquisition, and digital servicing. Lower expected lifetime returns reduce incentives to invest. Banks will respond by tightening underwriting, lowering credit limits, scaling back acquisitions, and reducing exposure to cyclical segments. Over time, these adjustments contract effective supply—even if accounting ROA never turns negative.

Price controls also distort risk across the business cycle. In expansions, upside is capped; in downturns, losses are not. This asymmetric payoff further discourages capital allocation to capped products. Banks earn less in good times while bearing full downside risk in bad times. The rational response is to shift capital toward assets where pricing can adjust with risk—a pattern well documented in the banking literature.[34]

In short, a positive projected ROA under a rate cap does not establish economic viability. The relevant test is whether returns exceed the full opportunity cost of capital on a risk-adjusted basis. The Vanderbilt study does not perform that test. By treating positive accounting returns as sufficient, it understates the likelihood of capital reallocation and credit contraction in the very segments the policy aims to help.

IX. Conclusion

The Vanderbilt Policy Accelerator study offers an ambitious attempt to quantify the effects of federal credit card interest-rate caps. Its central claim is that banks’ existing profit margins can absorb a mandated reduction in interest spreads while remaining profitable, thereby delivering net savings to consumers. That conclusion rests on a narrow comparative-static framework that asks only whether credit card lending remains arithmetically viable under a capped rate. It does not ask how market participants will respond once risk-based pricing is constrained.

Credit markets are dynamic. They rely on risk differentiation, portfolio optimization, and capital allocation under regulatory constraints. Theory and empirical evidence show that when a binding price ceiling prevents lenders from pricing for risk, adjustment occurs along other margins. Credit supply contracts for higher-risk borrowers, shifts toward lower-risk borrowers, and migrates into substitute or informal channels. Lenders restructure revenue through fees and less-salient pricing terms. Capital flows toward products with stronger risk-adjusted returns.

The Vanderbilt framework cannot capture these dynamics. By treating positive projected ROA as evidence of continued supply, it overlooks the relevant benchmark: risk-adjusted return on capital. Credit card lending is unsecured, operationally intensive, and procyclical. When caps compress margins, they reduce the risk premium available to compensate for volatility and downturn exposure. Even if accounting returns remain positive, capital may rationally move elsewhere.

The study’s uniform-pricing assumption further undermines its conclusions. If issuers continue to price below the cap for prime borrowers, the study overstates consumer savings. If they instead converge on the cap, the policy risks functioning as a price floor for low-risk borrowers. Either outcome calls the projected savings into question.

Most importantly, the study evaluates welfare only for borrowers who remain in the capped market. It does not account for those excluded from credit, those pushed into higher-cost or less-regulated alternatives, reduced innovation, or long-run contractions in supply. Transfers within a shrinking market do not constitute net welfare gains.

Price ceilings in credit markets do not eliminate risk or create credit. They reallocate it. They change who receives credit, on what terms, and from which providers. Whether that reallocation improves welfare depends on general-equilibrium effects—not on whether a mandated rate aligns with current accounting margins.

The relevant policy question is not whether banks can absorb lower interest spreads in a static model. It is whether restricting risk-based pricing improves consumer welfare once rationing, substitution, fee restructuring, and capital reallocation are taken into account. On that question, the weight of theory and evidence is far less reassuring than the Vanderbilt study suggests.

[1] 10 Percent Credit Card Interest Rate Cap Act, S. 381, 119th Cong. (2025), https://www.congress.gov/bill/119th-congress/senate-bill/381.

[2] Bernie Sanders, News: Sanders, Hawley Introduce Bill Capping Credit Card Interest Rates at 10%, U.S. Senate (Feb. 4, 2025), Press Release, https://www.sanders.senate.gov/press-releases/news-sanders-hawley-introduce-bill-capping-credit-card-interest-rates-at-10.

[3] Brian Shearer, Capping Credit Card Rates, Vanderbilt Policy Accelerator (Sept. 2025), https://cdn.vanderbilt.edu/vu-URL/wp-content/uploads/sites/412/2025/10/01144344/Capping-Credit-Card-Rates.pdf [hereinafter Vanderbilt Study].

[4] Itamar Drechsler et al., Credit Card Banking, Fed. Rsrv. Bank of N.Y. Staff Rep. No. 1143 (2025), https://www.newyorkfed.org/medialibrary/media/research/staff_reports/sr1143.pdf [hereinafter Drechsler Report].

[5] Vanderbilt Study, supra note 3, at 7, 17.

[6] Joseph E. Stiglitz & Andrew Weiss, Credit Rationing in Markets with Imperfect Information, 71 Am. Econ. Rev. 393 (1981).

[7] Vanderbilt Study, supra note 3, at 24.

[8] Id. at 25.

[9] Rajashri Chakrabarti et al., Less for You, More for Me: Credit Reallocation and Rationing Under Usury Limits, Fed. Rsrv. Bank of N.Y. Staff Rep. No. 1173 (2025).

[10] Samuel Munzele Maimbo & Claudia Alejandra Henriquez Gallegos, Interest Rate Caps Around the World: Still Popular, But a Blunt Instrument, World Bank Pol’y Rsch. Working Paper No. 7070 (2014).

[11] Consumer Fin. Prot. Bureau, Consumer Use of Buy Now, Pay Later and Other Unsecured Debt 1 (Jan. 2025).

[12] Id. at 14.

[13] Shigeru Sato & Shingo Kawamoto, Loan-Shark Lending Surge Feared in Japan, Bloomberg (Aug. 8, 2012), https://www.bloomberg.com/news/articles/2012-08-07/loan-shark-lending-surge-feared-in-japan.

[14] Aurora Ferrari et al., Interest Rate Caps: The Theory and the Practice, World Bank Pol’y Rsch. Working Paper No. 8398, at 18–20 (2018).

[15] Sumit Agarwal et al., Regulating Consumer Financial Products: Evidence from Credit Cards, 130 Q.J. Econ. 111, 112 (2015).

[16] Consumer Fin. Prot. Bureau, CARD Act Report 25 (2013), https://files.consumerfinance.gov/f/201309_cfpb_card-act-report.pdf.

[17] Consumer Bankers Ass’n, Facts Matter: CARD Act Report Reveals Credit Card Fee Landscape in Stark Contrast to CFPB’s Misleading Headlines (2024), https://consumerbankers.com/press-release/facts-matter-card-act-report-reveals-credit-card-fee-landscape-in-stark-contrast-to-cfpbs-misleading-headlines.

[18] Agarwal et al., supra note 15, at 157–58.

[19] Scott T. Nelson, Private Information and Price Regulation in the US Credit Card Market, 93 Econometrica 1371, 1371 (2025).

[20] Id. at 1372.

[21] Yiwei Dou et al., The Credit Card Act and Consumer Debt Structure, 7 J.L. Fin. & Acct. 91, 91-92 (2022).

[22] Song Han, Benjamin J. Keys & Geng Li, Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings, Bd. of Governors of the Fed. Rsrv. Sys., Fin. & Econ. Discussion Series No. 2015-103 (2015).

[23] Gregory Elliehausen & Simona M. Hannon, The Credit Card Act and Consumer Finance Company Lending, 34 J. Fin. Intermediation 109, 109-10 (2018).

[24] Vanderbilt Study, supra note 3, at 14 (“[P]ayday lenders uniformly charge the maximum interest rate allowed by state law.”).

[25] Id. at 17-18.

[26] Drechsler Report, supra note 4, at 29 (noting that Capital One’s marketing accounts for about 20% of total expenses); Vanderbilt Study, supra note 3, at 18 (assuming 20% of operating expenses in each tier go to advertising based on that estimate).

[27] Joel M. Bessis, Risk Management in Banking 197–205 (4th ed. 2015) (explaining RAROC frameworks, internal capital allocation, and risk-adjusted performance measurement in banking institutions).

[28] Ronald J. Mann, Charging Ahead: The Growth and Regulation of Payment Card Markets 96–101 (2006) (describing cyclicality and loss volatility in credit card portfolios); Fed. Deposit Ins. Corp., Quarterly Banking Profile tbl. RC-N (various years) (showing elevated, recession-sensitive net charge-off rates for credit card loans relative to other loan categories).

[29] Basel Comm. on Banking Supervision, Basel III: A Global Regulatory Framework for More Resilient Banks and Banking Systems ¶¶ 49–54, 153–159 (rev. June 2011) (describing capital buffers and the treatment of credit risk exposures); 12 C.F.R. pt. 3, subpt. D (2024) (OCC risk-based capital rules); 12 C.F.R. pt. 217, subpt. D (2024) (Federal Reserve capital rules implementing Basel standards).

[30] Basel Comm. on Banking Supervision, International Convergence of Capital Measurement and Capital Standards ¶¶ 330–335 (June 2006) (assigning risk weights and capital treatment for qualifying revolving retail exposures, including credit card receivables).

[31] Basel Comm. on Banking Supervision, Basel III: The Liquidity Coverage Ratio and Liquidity Risk Monitoring Tools ¶¶ 14–25 (Jan. 2013) (establishing liquidity coverage requirements); 12 C.F.R. pt. 249 (2024) (U.S. liquidity coverage ratio rule).

[32] Robert C. Merton & André F. Perold, Theory of Risk Capital in Financial Firms, 1 J. Applied Corp. Fin. 16, 19–24 (1993) (developing internal risk-capital allocation principles and hurdle-rate logic for financial institutions).

[33] Franklin Allen & Douglas Gale, Competition and Financial Stability, 55 J. Money, Credit & Banking 453, 458–63 (2004) (analyzing capital-allocation decisions and risk-taking incentives under competitive constraints).

[34] Anil K. Kashyap & Jeremy C. Stein, Cyclical Implications of the Basel II Capital Standards, 49 Econ. Persp. 18, 20–24 (2004) (discussing the procyclicality of bank capital requirements and lending behavior); Viral V. Acharya et al., Capital Shortfall: A New Approach to Ranking and Regulating Systemic Risks, 102 Am. Econ. Rev. 59, 61–64 (2012) (analyzing capital adequacy and downside risk exposure in financial institutions).

Executive Summary

Several states are considering legislation that would require large grocery suppliers to offer identical prices and contract terms to all retailers purchasing the same quantity of goods. New York Senate Bill S8563—the “Consumer Grocery Pricing Fairness Act”—is a leading example.

The proposal aims to prevent large national retailers from negotiating better wholesale prices than smaller independent grocers. Supporters argue that large retailers use their buying power to secure discounts unavailable to smaller stores, creating a “waterbed effect” in which suppliers recover lost revenue by charging smaller retailers higher prices. They contend that this dynamic harms independent grocers and contributes to food deserts in low-income communities.

This issue brief concludes that, while these concerns are understandable, the proposed legislation would likely raise grocery prices, reduce competition, and harm the consumers it intends to help.

S8563 would require large food manufacturers—those with more than $6 billion in annual sales—to offer identical “terms of sale” to retailers purchasing the same volume of goods, defined as up to a single truckload. “Terms of sale” include prices, discounts, rebates, delivery terms, promotional allowances, and other contractual provisions. The bill would also require suppliers to disclose anonymized contract terms offered to dominant retailers upon request. Violations could trigger enforcement actions by the state attorney general or private plaintiffs, with penalties of up to three times actual damages.

In practice, the legislation establishes a mandatory price-parity rule. If a supplier offers a discount to one covered retailer, it must extend the same terms to all similarly situated buyers. Because any discount would immediately apply across the market, suppliers would have strong incentives to eliminate targeted discounts and adopt uniform, higher wholesale prices.

Price differences across buyers are common in competitive markets and often reflect legitimate efficiencies. Large retailers frequently receive lower wholesale prices because they purchase in higher volumes, place predictable orders, and reduce suppliers’ distribution and marketing costs. Those efficiencies lower suppliers’ per-unit costs and often translate into lower retail prices for consumers. Research shows that large-format retailers such as Walmart frequently offer lower grocery prices and that their presence can prompt competing stores to reduce prices as well.

Mandatory price uniformity weakens those competitive dynamics. Economic theory and historical experience—particularly under the Robinson-Patman Act—suggest that price-parity rules discourage discounting and push suppliers toward rigid, uniform pricing structures.

Supporters of the legislation rely heavily on the “waterbed effect,” the idea that discounts to large buyers force suppliers to charge higher prices to smaller ones. While this concept appears in economic theory, it arises only under specific market conditions and has limited empirical support. Suppliers typically set prices based on each buyer’s demand conditions and competitive alternatives, not to offset discounts offered to other customers.

The bill’s disclosure requirements raise additional concerns. By requiring suppliers to disclose anonymized contract terms from dominant retailers, the legislation could reduce incentives to negotiate discounts. Economic research shows that greater pricing transparency in concentrated markets can facilitate coordination and weaken price competition. Advances in artificial intelligence further complicate the assumption that disclosed contract terms can remain effectively anonymized.

The legislation is also justified as a response to food deserts and nutritional inequality. Empirical research, however, indicates that limited geographic access to grocery stores explains only a modest portion of observed differences in diet quality. Much of the disparity reflects differences in consumer demand, rather than wholesale pricing conditions.

If the legislation raises wholesale prices or reduces discounting, the resulting retail price increases are likely to fall disproportionately on low-income households that rely on high-volume, low-margin retail formats. Increased regulatory risk could also discourage investment in new stores, particularly in underserved communities.

Policies aimed at supporting small grocers and improving food access address legitimate concerns. Mandatory wholesale price uniformity, however, is a blunt instrument. The available economic evidence suggests that such legislation would reduce discounting, increase grocery prices, and ultimately harm consumers—particularly those with the fewest alternatives.

I. Introduction and Overview

New York Senate Bill S8563—the “Consumer Grocery Pricing Fairness Act”—would prohibit large grocery suppliers from offering different prices, rebates, promotional allowances, or contract terms to retailers purchasing equivalent quantities of the same goods.[1] The bill applies to suppliers with more than $6 billion in annual sales and imposes additional restrictions on retailers with more than $18 billion in annual sales operating in more than 20 states. Violations would carry civil penalties of up to three times actual damages, enforceable by the New York attorney general or by private plaintiffs. Several other states are considering similar legislation.

Sen. Cordell Cleare, the bill’s sponsor, argues that large retailers extract preferential wholesale terms from suppliers.[2] Suppliers then raise prices to smaller, independent grocers to offset those discounts, creating what she describes as a “waterbed effect.” In her view, this price discrimination drives independent grocers out of business and contributes to the emergence of “food deserts,” where low-income neighborhoods lack adequate access to grocery stores.

This issue brief evaluates those claims against the relevant economic literature. It concludes that S8563—and similar bills under consideration in other states—would likely produce the opposite of their stated objectives. By prohibiting the negotiated discounts through which competitive pressure operates in grocery supply chains, the bill would reduce discounting across the market, raise average wholesale costs, and ultimately increase retail prices. The burden would fall most heavily on low-income consumers who rely on high-volume, low-margin retail formats.

The bill’s food-desert rationale also rests on a misreading of the empirical literature on nutritional inequality. Its mandatory disclosure provisions introduce an additional risk by facilitating the sort of coordinated pricing behavior that antitrust law seeks to prevent.

S8563 closely resembles pending legislation in Minnesota, Arizona, and Rhode Island, suggesting a coordinated multistate effort to establish a de facto national pricing standard through state law. Each proposal shares the same structural flaw: they treat price differences as presumptively anticompetitive without requiring any showing of consumer harm. That approach inverts the core logic of modern antitrust analysis.

II. How the Bill Works

S8563 would add a new Article 22-C to New York’s General Business Law. Its core provision—Section 350-k—makes it unlawful for a “covered supplier” to fail to extend the same “terms of sale” to all buyers purchasing on the same “volume unit basis” in “reasonably contemporaneous sales.” The bill defines “terms of sale” broadly. It covers not only price, but also discounts, rebates, delivery terms, payment terms, package size, promotional allowances, marketing devices, merchandising arrangements, and distribution terms, all evaluated on a per-unit basis.

This structure differs sharply from contractual most-favored-nation (MFN) clauses. In ordinary commercial practice, an MFN clause is a negotiated contractual provision under which a seller agrees to give a particular buyer terms no worse than those offered to other buyers. S8563 would effectively impose a statutory MFN regime. Covered suppliers would be legally required to offer the same terms to every similarly situated retailer, regardless of the parties’ negotiated agreements.

The bill defines “volume unit basis” as a base unit of measurement that cannot exceed a single truckload. That definition has significant practical consequences. The law treats the first truckload sold to an independent grocer the same as the 500th truckload sold to a national chain. Suppliers therefore must offer identical per-unit terms to both buyers. They cannot price the difference in predictability, logistics simplicity, or long-term planning value that a 500-truckload relationship provides.

TABLE 1: Major Retailers Covered by State Bills

[3] [4] [5] [6] [7]

The bill’s thresholds target the largest participants in the national grocery market while nominally exempting smaller firms. A “covered supplier” is any entity that produces and sells “covered goods” in New York with annual sales exceeding $6 billion (adjusted for inflation).[8] A “dominant covered retailer” is a retailer with more than $18 billion in annual retail sales of covered goods and at least one storefront or distribution center in more than 20 states, including New York.

The $18 billion retailer threshold captures Walmart, Costco, Ahold Delhaize, Whole Foods, Dollar General, and other national chains (Table 1). The $6 billion supplier threshold captures nearly all major branded food manufacturers—companies like PepsiCo, Nestlé, Kraft Heinz, General Mills, Conagra Brands, and Tyson Foods.[9]

A. Mandatory Disclosure of Competitors’ Pricing Terms

Section 350-k(2) requires that, upon written request from any covered retailer or wholesaler, a covered supplier must provide—within 14 days—anonymized terms of sale from all contracts with dominant covered retailers for the same covered good and volume unit basis during the prior 180 days. The provision converts competitively sensitive pricing information from a private bargaining asset into a quasi-public document accessible to any buyer that asks for it.[10]

B. Conditional Limits on Suppliers’ Ability to Refuse to Deal

The bill prohibits covered suppliers from refusing to sell to a non-dominant covered retailer that has paid on time during the previous 12 months and has requested equal terms, unless the supplier can demonstrate a “commercially reasonable justification.” The bill includes a construction clause stating it does not “require” doing business with anyone. In practice, however, the commercially reasonable justification standard functions as a conditional duty to deal. That requirement departs from the general rule in U.S. antitrust law that firms retain broad discretion to choose their trading partners.

C. Enforcement and Treble-Damages Liability

The New York attorney general may bring enforcement actions, and any “injured person” may file a private lawsuit. Available remedies include injunctions and civil penalties of up to three times actual damages—a treble-damages structure that mirrors the private remedy available under federal antitrust law. The bill also imposes agency liability for violations committed by contracted third parties, meaning a supplier cannot avoid liability by routing transactions through intermediaries.

The bill includes several affirmative defenses. Suppliers may justify pricing differences that are “predominantly attributable” to genuine distribution or manufacturing efficiencies, distress sales of perishable or seasonal goods, or coercion by a dominant retailer. The coercion defense requires the supplier to demonstrate that it did not collude, would have suffered substantial harm by refusing the dominant retailer’s demands, and made a good-faith effort to report the conduct to the attorney general—effectively requiring suppliers to act as state informants against their largest customers.

III. Parallel Legislation in Other States

S8563 forms part of a broader multistate legislative effort built on nearly identical statutory language, definitions, and thresholds.

A. Minnesota House File 2149 and Senate File 2556

HF2149 was introduced in the Minnesota House of Representatives March 10, 2025, and referred to the Committee on Commerce Finance and Policy.[11] Its Senate companion, SF2556, was referred to the Committee on Commerce and Consumer Protection March 13, 2025.[12] Both bills replicate S8563’s core structure: a “same terms of sale” mandate, a 14-day disclosure window, a 180-day lookback period, and a treble-damages enforcement regime. Reporting in Minnesota House Session Daily describes supporters arguing that grocery-market concentration has contributed to “food deserts expanding in rural and underserved areas.”[13]

B. Arizona Senate Bill 1226 and House Bill 2948

SB1226 and HB2948 were introduced in the Arizona Legislature in 2026.[14] Both adopt the same $6 billion supplier threshold and $18 billion retailer threshold. They also replicate the bill’s “volume unit basis” definition—capped at a single truckload—and the 14-day anonymized disclosure requirement. In their operative provisions, the Arizona bills are functionally identical to S8563.

C. Rhode Island House Bill 7514

Introduced Feb. 6, 2026, and referred to the House Committee on Corporations, H7514 follows the same legislative template: a same-terms mandate, anonymized disclosure of dominant-retailer contract terms, and enforcement by the state attorney general and private plaintiffs.[15]

D. Maine LD1894/HP1265

The Maine Legislature has considered related legislation.[16] In the current session, however, the bill’s title reflects a conversion to a resolution establishing a study commission, rather than imposing a same-terms-of-sale mandate.[17] Maine’s proposal remains conceptually related, but it does not currently replicate S8563’s operative structure.

The near-uniformity of definitions, thresholds, and enforcement mechanisms across these bills strongly suggests coordinated model-legislation drafting. Multistate enactment would create effects that extend beyond any single jurisdiction. A supplier operating in New York, Minnesota, Arizona, and Rhode Island would face substantial compliance risk if it offered differentiated terms to retailers in any one of those states. The most likely response would not be state-by-state compliance. Suppliers would instead adopt uniform national pricing—eliminating the competitive discounting that currently benefits consumers across all states, not just those that enact the legislation.

IV. The Robinson-Patman Act and S8563

The Robinson-Patman Act (RPA) of 1936, codified at 15 U.S.C. § 13, prohibits sellers engaged in interstate commerce from charging competing buyers different prices for goods of “like grade and quality” where the effect “may be substantially to lessen competition or tend to create a monopoly.” The statute also covers discriminatory promotional allowances and services under Sections 2(d) and 2(e). The Federal Trade Commission (FTC) administers the act and publishes compliance guidance, including the Fred Meyer Guides governing promotional allowances, codified at 16 C.F.R. Part 240.

Congress enacted the RPA in response to the competitive pressure that national grocery chains—particularly the Great Atlantic & Pacific Tea Company (A&P)—placed on smaller, independent grocers during the 1930s.[18] Legislators worried that A&P’s buying power allowed it to secure supplier discounts unavailable to smaller rivals, giving it a retail price advantage that independent stores could not match.

A. Modern Robinson-Patman Doctrine and Enforcement

Federal enforcement of the RPA declined sharply beginning in the late 1970s. The U.S. Department of Justice (DOJ) stopped bringing RPA cases in 1977, and the FTC largely ceased enforcement by the 1990s.[19] The Antitrust Modernization Commission, a bipartisan body established by Congress in 2002, issued a report in 2007 recommending outright repeal of the statute on the grounds that it discourages price discounting, raises consumer prices, and protects competitors, rather than competition:

By broadly discouraging price discounts, the Robinson-Patman Act potentially harms competition and consumers… the Act protects competitors, often at the expense of competition… All of these costs are likely to result in higher prices to consumers than would be the case if the Robinson-Patman Act were not on the books.[20]

The Supreme Court has interpreted the RPA in light of broader antitrust principles. In Brooke Group Ltd. v. Brown & Williamson Tobacco Corp., the Court held that the act condemns price discrimination only when it threatens competitive injury.[21] In Volvo Trucks N. Am., Inc. v. Reeder-Simco GMC, Inc., the Court further held that plaintiffs must demonstrate harm to competition among purchasers who actually compete with one another.[22]

Recent enforcement activity suggests a tentative revival of the statute. In January 2025, the FTC filed suit against PepsiCo under the RPA, alleging that the company offered Walmart promotional pricing and marketing support unavailable to smaller retailers.[23] The FTC dismissed the case without prejudice in May 2025.[24] Similarly, in FTC v. S. Glazer’s Wine & Spirits, LLC, the agency alleged that the distributor violated the RPA by offering large chains discounts unavailable to smaller retailers.[25] A district court denied the defendant’s motion to dismiss in April 2025,[26] although the case continues to face challenges, including potential cost-justification defenses.

Private enforcement persists as well. In February 2026, the 9th U.S. Circuit Court of Appeals affirmed a district court judgment for wholesale purchasers in LA Int’l Corp. v. Prestige Brands Holdings, Inc., involving promotional payment discrimination favoring Costco, while remanding on attorney’s fees.[27] Taken together, these developments suggest a modest resurgence of Robinson-Patman enforcement through both FTC activity and private litigation.

B. How S8563 Goes Beyond the Robinson-Patman Act

S8563 and similar state bills share the RPA’s core premise: that large buyers extract supplier concessions unavailable to smaller competitors, creating cost disadvantages that affect downstream retail competition. Like the RPA, the state-level “grocery pricing fairness” bills extend beyond list prices to cover promotional allowances, marketing devices, and merchandising arrangements—the same conduct addressed by Sections 2(d) and 2(e) of the RPA and the Fred Meyer Guides. The state proposals also incorporate familiar affirmative defenses, including cost justification and meeting competition, placing the burden on defendants to show that pricing differences reflect legitimate business reasons.

Despite these similarities, S8563 departs from the RPA in several important ways that make it substantially more restrictive.

First, the bill eliminates the competitive-injury requirement that limits RPA liability. The federal statute applies only when a pricing difference “may be substantially to lessen competition or tend to create a monopoly.” Courts interpreting this language require plaintiffs to demonstrate actual competitive harm; price differences alone do not suffice. S8563 imposes no such screening mechanism. A covered supplier violates the statute whenever it fails to offer identical terms to similarly situated buyers, regardless of whether the difference produces any competitive harm at the retail level. This categorical rule removes the central constraint that courts have used to prevent the RPA from condemning efficient, pro-competitive pricing.

Second, the bill’s truckload cap on the “volume unit basis” definition diverges sharply from established RPA doctrine. Under federal law, suppliers may offer volume discounts that reflect genuine economies of scale—lower per-unit costs resulting from larger, more predictable orders, simplified logistics, or reduced inventory risk. Courts and the FTC recognize these efficiencies as legitimate cost justifications.[28] S8563 forecloses that flexibility. Because the volume unit basis cannot exceed a single truckload, the law requires suppliers to offer the same per-unit terms to a buyer ordering one truckload as to a buyer ordering hundreds. A supplier that processes 500 weekly truckload orders from a national retailer—allowing it to plan production runs, reduce waste, and minimize carrying costs—must offer the same per-unit price to a local grocer placing irregular orders, regardless of the difference in planning value between the two relationships.

Third, the bill imposes a disclosure obligation that has no counterpart in the RPA. Federal law does not give buyers the right to compel disclosure of the terms a supplier offers to other customers. S8563’s 14-day disclosure requirement would convert competitively sensitive pricing information into a quasi-public document. The economic implications of this mandate are discussed in Section VI.

Fourth, S8563 introduces a conditional duty-to-deal rule that extends beyond the RPA’s framework. The RPA regulates discriminatory pricing, but it does not generally require suppliers to sell to any particular buyer. By contrast, S8563 prohibits suppliers from refusing to sell to certain non-dominant retailers unless they can demonstrate a “commercially reasonable justification.” In practice, this provision creates a de facto obligation to deal. U.S. antitrust law has long treated unilateral refusals to deal as legitimate business decisions, subject to narrow exceptions, as recognized in Aspen Skiing Co. v. Aspen Highlands Skiing Corp.[29] S8563 significantly narrows that discretion.

Finally, the bill applies only to firms above specified revenue thresholds—$6 billion for suppliers and $18 billion for retailers. The RPA contains no such size limitations; it applies to all sellers engaged in interstate commerce. By targeting firms based on scale alone, S8563 reflects a “big is bad” approach that sits uneasily with the competitive-effects analysis that governs modern antitrust law.

V. The ‘Waterbed Effect’: Theory, Evidence, and Misapplication

The central economic justification offered for S8563 rests on the “waterbed effect”—the theory that price concessions to large buyers force suppliers to raise prices for smaller buyers. The sponsor’s memo asserts that “suppliers make up lost profits from discounted sales to dominant buyers by charging higher prices to smaller buyers,” contributing to the affordability pressures facing independent grocers.[30]

The metaphor derives from the behavior of a water-filled mattress: pressing down in one place causes another part to rise. Applied to retail markets, the concept suggests that lowering prices for one group of buyers may cause prices to increase for others.

A. The Waterbed Effect in Economic Theory

The modern formulation of the waterbed effect in retail markets originates with Roman Inderst and Tommaso Valletti. [31] Their 2011 paper provides the primary theoretical framework for the claim. [32]

Their model describes a market in which a large buyer negotiates a lower wholesale price from a supplier. The lower input cost allows the large buyer to reduce its retail price and capture market share from smaller rivals. As smaller retailers lose volume, their bargaining power with the supplier weakens. Because lower sales volume reduces the supplier’s incentive to maintain favorable terms, the supplier may charge those smaller buyers higher wholesale prices. Those higher input costs may then be passed through to consumers shopping at smaller stores.

Inderst and Valletti also identify the conditions under which a waterbed effect could harm consumers overall. Consumer harm occurs only if the price increase paid by customers of smaller retailers exceeds the price decrease enjoyed by customers of the large retailer. That outcome requires two conditions:

1. smaller retailers must account for a substantial share of the market, and
2. the large retailer must pass through its wholesale savings to consumers at a relatively low rate.

Absent those conditions, the price reductions offered by large retailers dominate the effect of any higher prices faced by smaller stores.

B. Antitrust Courts Treat Waterbed Claims Skeptically

Courts have generally treated waterbed claims with skepticism. In DeHoog v. Anheuser-Busch InBev, plaintiffs argued that Anheuser-Busch InBev’s acquisition of SABMiller would produce a waterbed effect in the market for beer ingredients.[33]

Under the plaintiffs’ theory, the newly enlarged ABI would gain substantial bargaining power over suppliers of hops, a key input in brewing. Suppliers forced to lower prices for ABI would supposedly recoup those losses by charging higher prices to smaller brewers, including craft breweries.

The court rejected the theory as speculative. The judge noted that antitrust law primarily protects the suppliers who might be squeezed by a large buyer, not the buyer’s competitors. The court also observed that the claim was internally inconsistent. If suppliers truly faced persistent price pressure from large brewers, they could simply shift production toward more profitable varieties of hops rather than accept lower prices indefinitely.

A similar claim arising in grocery markets would likely face the same doctrinal obstacles. Antitrust law addresses monopsony concerns primarily to protect suppliers, not competing retailers. Absent clear evidence of consumer harm, allegations that wholesale prices may shift across buyers would likely be treated as conjectural.

C. Empirical Evidence from Competition Authorities

Empirical evidence for a waterbed effect in retail markets remains limited. The United Kingdom’s competition authorities have repeatedly examined the theory in grocery-sector investigations and found little support for it.

In its 2003 investigation of the Safeway merger, the U.K. Competition Commission concluded there was “little evidence” of an immediate waterbed effect and insufficient data to suggest that further consolidation would exacerbate price disadvantages for smaller retailers.

Overall, therefore, there is little evidence of an immediate or short-term “waterbed” effect. … [O]ur surveys produced insufficient evidence on this point for us to conclude that any waterbed effect would be exacerbated by any of the mergers.[34]

Similarly, the Office of Fair Trading reported in 2006 that significant theoretical questions remained unresolved. These included why suppliers would persistently price below cost for large supermarkets or how they could charge smaller retailers above-market prices without being undercut by competing suppliers. [35] UK authorities ultimately viewed the waterbed effect as a theory lacking both a robust empirical foundation and practical applicability to the complex dynamics of grocery retail markets.

D. Why the Sponsor Memo Misapplies the Theory

The S8563 sponsor’s memo appears to conflate two distinct arguments.

The first is the Inderst–Valletti equilibrium model, which predicts a waterbed effect only under specific market conditions. The second is a simpler “subsidy” claim—that suppliers must raise prices for smaller buyers in order to recoup losses from discounts granted to larger buyers.

The subsidy claim is not consistent with standard microeconomic theory. Suppliers set prices for each buyer based on the demand conditions and competitive alternatives associated with that particular relationship. If a supplier could profitably charge higher prices to small grocers, profit-maximizing behavior predicts it would already do so—regardless of the terms offered to large retailers such as Costco, Walmart, or Target.

A discount granted to one buyer does not create a “gap” that must be filled by charging higher prices to another. Bruce Kobayashi and Timothy Muris describe this subsidy framing as an “incoherent notion” that misunderstands how profit maximization operates in markets with heterogeneous buyers.[36] Suppliers facing buyers with different demand elasticities naturally engage in price discrimination. The resulting pattern of higher and lower prices reflects differences in demand conditions, not cross-subsidization between buyers.

While the Inderst–Valletti model identifies conditions under which a waterbed effect could harm consumers, the drafters of S8563 make no attempt to assess whether those conditions exist in New York grocery markets. The legislation cites no empirical evidence regarding the market share of independent grocers, the rate at which large retailers pass wholesale savings through to consumers, or the magnitude of wholesale price differences attributable to bargaining power, rather than cost.

Absent that analysis, the waterbed theory functions less as an empirical finding than as a rhetorical justification for the legislation.

VI. The Disclosure Mandate and the Risk of Coordination

Section 350-k(2) of S8563 requires covered suppliers to disclose contract terms to requesting buyers. Within 14 days of a written request, a supplier must provide anonymized terms of sale from all contracts with dominant covered retailers for the same covered good and volume unit basis during the prior 180 days. The provision operates as a transparency rule. Armed with information about what Target, Stop & Shop (Ahold Delhaize), or Acme Markets (Albertsons) paid for a truckload of branded cereal, a regional independent grocer can demand equivalent terms.

A. The Bill’s Disclosure Requirement

The disclosure rule converts competitively sensitive pricing information into a shared market reference point. A supplier must reveal the terms offered to dominant retailers whenever another covered retailer asks for them. The bill assumes that these disclosures can be effectively anonymized.

That assumption may be unrealistic. Recent research shows that large language models can reidentify individuals or entities from ostensibly anonymized text by extracting identity-relevant features and matching them with external data sources.[37] With only a small number of retailers subject to the bill’s disclosure requirements, the risk of reidentification increases substantially. Even if supplier contracts omit retailer names, market participants may still infer which chain received which terms.

B. Transparency, Information Exchange, and Coordination Risk

The economics literature has long recognized that increased price transparency can weaken competition. George Stigler’s classic paper, “A Theory of Oligopoly,” shows that sustained coordination among firms depends on the ability to detect and punish deviations from a shared pricing norm.[38] Greater transparency makes that monitoring easier. When firms can observe pricing behavior more clearly, the threat of retaliation becomes more credible and the incentive to undercut rivals declines.

Competition authorities have reached similar conclusions. The OECD Competition Committee’s 2010 roundtable on information exchanges among competitors concluded that enhanced transparency can facilitate collusion, particularly when pricing information becomes widely observable.[39] The risk depends on several factors, including market structure, the nature of the information exchanged, and the frequency of disclosure. In its submission to the OECD roundtable, the United States likewise observed that, while information sharing may generate efficiencies, exchanges involving pricing data can create coordination risks and therefore receive rule-of-reason scrutiny under U.S. antitrust law.[40]

C. The Disclosure Rule Might Suppress Competitive Discounting

S8563’s disclosure mechanism operates through a vertical channel—supplier to retailer—rather than the horizontal information exchanges typically discussed in antitrust literature. The underlying economic logic nonetheless applies. Only a small number of firms qualify as dominant covered retailers—roughly a dozen chains listed in Table 1. In such a concentrated environment, anonymized disclosures may still allow market participants to infer which retailer received which price.

Once disclosed, the price becomes a market benchmark. A regional grocer that receives anonymized data showing that a dominant retailer paid $X per unit for a particular product can demand that same price. Because the statute requires disclosure within 14 days and applies to all similarly situated buyers, any discount offered to one retailer effectively becomes available to every retailer that requests it.

The result is a practical floor on competitive discounting. The lowest price offered to any dominant retailer becomes the minimum price that other buyers can demand. Suppliers therefore have little incentive to negotiate deep discounts with any individual buyer, because those concessions cannot remain private.

Evidence from other markets illustrates a similar mechanism. Research on pay transparency in labor markets finds that when wages become widely observable, employers often respond by compressing wage differences rather than maintaining individualized negotiations. Zoë Cullen and Bobak Pakzad-Hurson find that increased wage transparency leads firms to offer lower average wages in order to avoid triggering costly renegotiations across workers.[41]

The same logic applies to wholesale grocery pricing. When discounts cannot remain private, the equilibrium response is fewer discounts.

VII. Pricing, Competition, and Consumer Welfare

The central economic problem with S8563 and similar same-terms mandates is that they transform competitive discounts into compulsory across-the-board price concessions. Under current law, a supplier may offer Walmart a 12% discount because Walmart’s scale, logistics capabilities, and purchasing predictability reduce the supplier’s costs. Bilateral negotiation—backed by Walmart’s credible ability to switch suppliers or expand private-label offerings—creates competitive pressure that ultimately benefits consumers. The resulting discount reflects firm-specific efficiencies and market bargaining dynamics.

S8563 eliminates that firm-specific character. If a supplier grants a 12% discount to Walmart on a truckload purchase, the supplier must extend that same discount to any covered buyer purchasing on the same “volume unit basis,” regardless of whether those buyers generate comparable cost savings or strategic value. The discount ceases to be a targeted response to a particular buyer’s efficiencies and instead becomes a legally mandated public price term.

Faced with that rule, suppliers cannot confine discounts to transactions that justify them economically. Any discount offered to one major buyer automatically propagates to others. The rational response is therefore to reduce or eliminate targeted discounts and move toward a higher uniform price. Rather than protecting small retailers, the statute risks suppressing the bargaining process that produces lower prices for consumers.

A. Why Same-Terms Mandates Differ from MFNs

This mechanism differs fundamentally from a most-favored-nation (MFN) clause. An MFN is a privately negotiated contractual provision that ties a seller’s pricing across particular sales channels or counterparties. It does not require a seller to offer a discount; it merely restricts selective price reductions once granted.

A same-terms mandate operates differently. It is a statutory, market-wide obligation imposed on suppliers regardless of their contractual preferences. It applies to all covered transactions, eliminates the ability to tailor pricing to buyer-specific efficiencies, and transforms negotiated concessions into mandatory system-wide pricing commitments.

In short, MFNs operate within individual contracts and limit how a seller may vary prices among specific counterparties. Same-terms mandates apply across the market and shape how suppliers set wholesale prices more broadly. By requiring that discounts offered to one buyer be extended to others, same-terms mandates alter suppliers’ incentives to offer negotiated discounts in the first place.

B. Evidence from Robinson-Patman Enforcement

Historical experience with the Robinson-Patman Act illustrates this dynamic. The Antitrust Modernization Commission’s 2007 report concluded that price-discrimination liability often discouraged firms from offering discounts and instead pushed them toward rigid, uniform pricing to avoid litigation risk.

It is difficult to know the frequency and amounts of price discounts and corresponding savings for consumers that the Robinson-Patman Act has deterred. … Nonetheless, anecdotal evidence and informed judgment based on economic theory suggests that the additional costs to consumers of seventy years of forgone discounts are likely substantial. The Act’s continued existence can discourage firms from taking procompetitive actions because doing so might lead to litigation asserting Robinson-Patman Act claims that, even were the litigation to be resolved in the company’s favor, would involve distractions, expenses, and risks that make the procompetitive course of action not worth the cost of pursuing it.[42]

The commission concluded that the statute’s chilling effect on discounting likely increased retail prices relative to a world without Robinson-Patman enforcement.

C. Who Bears the Cost?

The distributional consequences of reduced discounting are not neutral. Emek Basker’s review of the literature on Walmart’s role in grocery markets finds that Walmart’s grocery prices are, on average, about 10% lower than those of its competitors. The entry of a Walmart Supercenter typically prompts competing grocery stores to reduce their prices by roughly 1% to 3%.[43] Walmart’s presence therefore lowers prices not only for its own customers but also for consumers who shop at competing stores.[44]

Jerry Hausman and Ephraim Leibtag similarly estimate, using household panel data, that the expansion of large-format supercenters generates substantial consumer-welfare gains. Low-income households benefit disproportionately because they shop at price-sensitive formats more frequently than higher-income households.[45]

If S8563 leads suppliers to reduce or eliminate volume discounts in order to avoid market-wide matching obligations, the price advantage currently offered by high-volume retailers will narrow or disappear. The consumers most affected by that change will be those who rely on low-price retail formats and have the fewest alternatives—the same population the sponsor’s memo identifies as the bill’s intended beneficiaries.

D. Effects on Non-Price Competition

S8563’s definition of “terms of sale” extends beyond price. The statute covers marketing devices, merchandising arrangements, and distribution terms—the full range of non-price benefits that suppliers provide retailers as part of broader supply relationships.

These arrangements are not interchangeable across retailers. National chains operate sophisticated retail media networks—digital advertising platforms through which suppliers purchase targeted placements based on consumer purchase data. A retailer such as Walmart can offer suppliers advertising placements on its digital platform that a neighborhood grocery store with 500 weekly customers cannot replicate.

Requiring that these arrangements be offered on “proportionally equal terms” does not create equivalent access. Instead, it creates legal exposure for any supplier that offers specialized programs to dominant retailers. The predictable response is to withdraw those arrangements altogether or remove them from supplier contracts with dominant retailers to avoid disclosure and matching obligations.

Consumers bear the cost of that withdrawal. Retail media revenue helps subsidize grocery prices at large retailers by offsetting other operating costs. Eliminating those arrangements reduces the ability of retailers to pass those savings through to shoppers.

VIII.  Litigation Risk and Compliance Costs

S8563’s treble-damages enforcement structure exposes covered suppliers to substantial litigation risk. The statute authorizes enforcement actions by the New York attorney general and private plaintiffs. As a result, routine pricing decisions could become the basis for lawsuits alleging unlawful price discrimination.

This exposure is significant because the statute regulates not only list prices but also discounts, rebates, promotional allowances, merchandising arrangements, and other “terms of sale.” Each difference in pricing or contractual treatment across buyers could become a potential cause of action. The combination of broad substantive coverage and treble damages creates strong incentives for litigation.

A. The Difficulty of the Cost-Justification Defense

The statute includes a cost-justification defense that allows suppliers to defend pricing differences by showing that they are “predominantly attributable” to genuine cost efficiencies. In practice, however, this defense has proven difficult to apply.

As John Yun explains in the context of the Robinson-Patman Act:

[W]hile the [RPA] permits cost differences as a defense for differential pricing, reliably separating cost-justified differences from price discrimination is difficult. What if 80% of a price difference is explained by costs but 20% is not? Is that still a violation? This line-drawing problem is not trivial. It imposes (a) significant evidentiary and administrative burdens on firms, enforcers, and courts and (b) social costs related to false positives, given that quantity discounts are ubiquitous in a market economy.[46]

These evidentiary challenges create legal uncertainty even when pricing differences reflect legitimate efficiencies. Suppliers may struggle to demonstrate precisely how logistical, manufacturing, or administrative cost savings translate into specific price differences across transactions.

B. Operational Compliance Costs

The practical burden of complying with the statute would be substantial. Covered suppliers selling hundreds of products to dozens of retailers in New York would need to maintain detailed documentation explaining the cost basis for every pricing difference across buyers.

That documentation would need to track the logistical, manufacturing, and administrative cost drivers associated with each product, buyer, and transaction. Without such records, suppliers would face significant litigation risk in private treble-damages actions—even when their pricing decisions reflect legitimate economic considerations.

Maintaining this level of transaction-specific cost documentation would impose significant administrative costs. Those compliance expenses ultimately flow through supply chains and into retail prices, reducing any consumer benefits the legislation’s proponents anticipate.

IX. Food Deserts: Evidence and Policy Mismatch

S8563’s sponsor memo asserts that pricing practices by dominant retailers have forced independent grocers to close, leaving low-income neighborhoods without access to fresh, nutritious food. This claim links the bill to a body of public health and food-policy research on “food deserts”—areas where residents have limited access to affordable, nutritious food.

A. What ‘Food Deserts’ Are—and How Common They Are

The U.S. Department of Agriculture’s (USDA) Economic Research Service (ERS) defines low-income, low-access census tracts as areas where at least 500 residents, or 33% of the population, live more than one mile from a supermarket in urban areas or more than 10 miles from a supermarket in rural areas.[47] The ERS estimated in its 2009 Report to Congress that roughly 23.5 million Americans live in low-income, low-access areas. At the same time, only about 2.3 million households—roughly 10% of that population—both lack access to a vehicle and live more than one mile from a supermarket.[48]

These figures highlight an important point: while geographic access to grocery stores varies across communities, most households in low-income, low-access areas retain the ability to travel to supermarkets outside their immediate neighborhoods. Limited food access therefore reflects a combination of transportation constraints, residential patterns, and consumer behavior—not simply the absence of nearby stores.

B. What the Research Shows About the Causes of Food Deserts

The empirical literature finds that wholesale pricing differentials play only a limited role in determining where supermarkets locate or why nutritional inequality persists.

Hunt Allcott and his co-authors combine evidence from supermarket entry events and household moves across neighborhoods with a structural model of grocery demand to simulate the effects of equalizing access to stores and prices across income groups.[49] Their counterfactual simulation finds that giving low-income households the same grocery access and prices as higher-income households would reduce nutritional inequality—measured by the healthfulness of food purchases—by roughly 10%. The remaining 90% of the gap reflects differences in demand patterns across income levels.

This finding does not render food access irrelevant. It instead suggests that access constraints are a secondary driver of nutritional inequality. Policies that operate only through the access channel—such as legislation intended to improve independent grocers’ wholesale purchasing conditions—address a margin that explains only a small portion of the observed disparity. S8563 also operates even further upstream: it targets wholesale pricing practices, rather than retail access itself.

Steven Cummins, Ellen Flint, and Stephen Matthews provide additional evidence from a natural experiment involving the opening of a new supermarket in a Philadelphia food desert.[50] Using a controlled pre-post quasi-experimental longitudinal design, they find that the new store improved residents’ perceptions of food access. Six months after the store opened, however, residents’ fruit and vegetable consumption and body mass index showed no measurable change. Improvements in retail infrastructure alone therefore may not produce the expected changes in consumption patterns.

C. How S8563 Could Worsen Food Access

The mechanisms described in earlier sections suggest that S8563 could worsen the very access problems it aims to address. The same forces that raise retail prices under a mandatory parity regime—reduced supplier discounting, withdrawal of promotional support, and compliance-driven price standardization—raise costs for all retailers.

Stores operating on thin margins in low-income urban neighborhoods face particularly high operating costs, including security, spoilage, and insurance. These stores are therefore more sensitive to increases in wholesale prices than supermarkets serving higher-income, higher-volume suburban markets. If S8563 compresses the wholesale price advantages currently available to high-volume retailers, it simultaneously raises cost pressures on the marginal stores most likely to close—small independent grocers operating in underserved communities.

The legislation may also discourage the expansion of high-volume retail formats in New York. Hausman and Leibtag show that the entry of supercenters produces consumer-welfare gains that disproportionately benefit low-income households.[51] If national chains conclude that New York’s regulatory environment makes aggressive wholesale negotiation legally risky, the expected returns from opening new stores in the state decline. Communities with limited retail infrastructure are the most dependent on such entry.

Finally, suppliers responding to the bill’s same-terms mandates may simplify their pricing structures by offering fewer products in fewer configurations. Because complex, differentiated contracts create legal exposure, standardization reduces litigation risk. The result could be narrower product assortments across retail formats. A statutory parity mandate thus creates incentives for suppliers to reduce differentiation, sacrificing variety in order to minimize compliance risk.

X. Error Costs and Antitrust Policy

A central concern in law & economics analysis of antitrust policy is the cost of errors—specifically, the asymmetric consequences of condemning efficient conduct (a false positive, or Type I error) versus failing to condemn genuinely harmful conduct (a false negative, or Type II error). Antitrust rules that cast too wide a net risk deterring procompetitive behavior, while rules that are too permissive risk allowing genuinely anticompetitive practices to persist.

Because competitive markets rely on experimentation in pricing, contracting, and distribution, false positives tend to impose especially high costs. When lawful conduct is mistakenly condemned, firms often respond by abandoning practices that benefit consumers, even if those practices ultimately would have been upheld in court. The result is a chilling effect on competitive behavior.

A. The Risk of False Positives Under S8563

S8563 and similar legislation increase the risk of false positives by treating any difference in terms of sale between buyers on the same volume unit basis as presumptively unlawful. Under this framework, every negotiated discount, customized promotional arrangement, or logistics-adjusted rebate becomes a potential statutory violation—regardless of whether the pricing difference reflects legitimate cost savings, demand differences, or ordinary competitive bargaining.

The consequences of a false positive under the statute are substantial. A covered supplier found liable faces treble damages, injunctive relief requiring it to restructure its pricing across all buyers in the state, and the legal and compliance costs associated with defending the litigation. Even suppliers that ultimately prevail at trial must bear the cost and uncertainty of the process.

These incentives encourage defensive behavior. Firms may abandon competitive discounting practices simply to reduce litigation exposure. As discussed in Section VIII, the Antitrust Modernization Commission documented precisely this response in the context of Robinson-Patman enforcement. Firms often adopted rigid, uniform pricing rather than risk litigation over selective discounts. S8563’s private right of action and treble-damages structure create at least as strong an incentive for such defensive pricing strategies.

B. False Negatives Are Already Addressed Under Existing Law

The cost of a false negative in this setting—failing to detect a case in which a dominant retailer coerces a supplier into offering discriminatory terms that harm competition—is comparatively limited because existing legal tools already address such conduct.

The Robinson-Patman Act remains available to challenge discriminatory pricing practices that threaten competition. More broadly, anticompetitive uses of buyer power can be pursued under Section 2 of the Sherman Act or under Section 5 of the FTC Act. These statutes allow regulators and private plaintiffs to challenge conduct that produces genuine competitive harm without imposing a categorical ban on negotiated pricing differences.

The relevant policy question is therefore not whether any instance of buyer-power abuse might escape scrutiny under existing law. The relevant question is whether a categorical prohibition on price differences—one that eliminates the competitive-effects screen used in modern antitrust analysis—produces a better balance of error costs than targeted enforcement under existing statutes. The available evidence suggests that it does not.

XI. More Effective Policy Alternatives

The concerns motivating S8563 and similar bills—rising grocery prices, declining viability of independent grocers, and inadequate food access in low-income neighborhoods—are real. The legislation’s diagnosis of those problems, however, is incomplete. More targeted policy instruments exist to address each concern directly.

A. Targeted Antitrust Enforcement

Existing antitrust law already provides tools to address coercive conduct by dominant retailers. Practices such as credible threats to delist suppliers who refuse exclusionary terms, demands for payments lacking a plausible efficiency justification, or contractual arrangements designed to raise rivals’ costs can be challenged under current law.

The FTC, state attorneys general, and private plaintiffs all possess authority to bring such cases. Targeted enforcement against specific anticompetitive practices addresses genuine abuses of buyer power without imposing a categorical prohibition on negotiated pricing differences that often reflect legitimate efficiencies.

B. Direct Policies to Improve Food Access

Food access challenges in particular communities stem largely from transportation constraints and the costs of operating grocery stores in underserved neighborhoods. The USDA’s 2009 Report to Congress emphasizes the importance of these factors in shaping food access patterns.

Policies that directly address those constraints are more likely to improve outcomes. Zoning reforms that reduce barriers to grocery-store entry and transportation investments that connect residents to existing retail locations can improve access without distorting wholesale pricing relationships.

New York already operates programs designed to address these issues, including tax incentives and financing initiatives that encourage grocery investment in underserved areas. These targeted interventions operate directly on the factors that determine store entry and access, making them more effective tools than a statewide mandate governing supplier pricing practices.

C. Targeted Enforcement Against Harmful Information Exchanges

Anticompetitive information exchanges can also be addressed through existing antitrust tools. Competition authorities in the United States and abroad have identified specific practices that raise coordination concerns, including coordinated price announcements, exchanges of forward-looking pricing information, and hub-and-spoke arrangements that facilitate collusion.

Targeted enforcement against these practices addresses genuine risks to competition. By contrast, S8563 mandates broad disclosure of pricing information as a remedy. Rather than preventing harmful exchanges, the statute risks creating the very transparency conditions that can facilitate coordination among market participants.

XII. Conclusion

State-level grocery “pricing fairness” legislation attempts to address concerns about grocery affordability and food access by restricting how suppliers and retailers negotiate prices. The economic evidence, however, suggests these bills misunderstand how pricing works in competitive supply chains. Discounts offered to large retailers are not simply transfers that disadvantage smaller stores. They often reflect cost savings, bargaining dynamics, and operational efficiencies that translate into lower retail prices for consumers.

By requiring suppliers to offer identical terms across buyers, S8563 would likely discourage these discounts and push suppliers toward uniform, higher wholesale prices. The bill’s disclosure requirements and litigation risks could further weaken competitive negotiation and increase compliance costs. Those costs would ultimately be passed through the supply chain to consumers.

For these reasons, S8563 and similar proposals risk producing outcomes that run counter to their stated goals. Rather than protecting small grocers or improving food access, the policy could raise grocery prices, reduce product variety, and discourage investment in high-volume retail formats that currently deliver some of the lowest prices for low-income households.

Addressing concerns about market power, grocery affordability, and food deserts requires more targeted policy tools. Policymakers seeking to improve consumer outcomes would be better served by strengthening existing enforcement mechanisms and pursuing policies that directly address transportation, zoning, and investment barriers in underserved communities.

[1] S. 8563, 2025–2026 Reg. Sess. (N.Y. 2025), https://legislation.nysenate.gov/pdf/bills/2025/S8563.

[2] Sponsor’s Mem. (Cleare), Bill No. S. 8563 (N.Y. 2025), https://www.nysenate.gov/legislation/bills/2025/S8563 [hereinafter Sponsor’s Memo].

[3] Figures reflect the most recently reported fiscal year.

[4] Reflects Amazon’s ownership of Whole Foods Market stores.

[5] Amazon Staff, Amazon Doubles Down on Online Grocery Delivery and Whole Foods Market Expansion to Reach More Customers, Amazon News (Jan. 27, 2026), https://www.aboutamazon.com/news/company-news/amazon-fresh-go-stores-closing-expanding-whole-foods (“Amazon is one of the top three grocers in the United States, with over $150 billion in gross sales.”).

[6] Koninklijke Ahold Delhaize N.V., Annual Report 2025 (2026), https://cms.aholddelhaize.com/media/toibtcys/ahold-delhaize-annual-report-2025-interactive.pdf.

[7] See The PG 100: Annual Ranking of Top Food Retailers, Progressive Grocer (May 19, 2025), https://progressivegrocer.com/pg-100-annual-ranking-top-food-retailers. The ranking includes revenue from non-grocery sales, such as fuel and general merchandise.

[8] A “covered good” includes the typical inventory found in supermarkets. The bill defines the term broadly to include standard grocery items—specifically those classified as “eligible foods” under federal SNAP regulations, 7 C.F.R. § 271—as well as general consumer packaged goods. It excludes gasoline, prescription drugs, tobacco, and alcoholic beverages.

[9] Food Processing’s Top 100—2024, Food Processing (Aug. 2024), https://www.foodprocessing.com/top100/2024chart.

[10] The provision may raise First Amendment concerns, though a full analysis lies beyond the scope of this issue brief. See, e.g., Petition for Writ of Certiorari, PhRMA v. O’Day, No. 25-1018 (U.S. Feb. 20, 2026), https://www.supremecourt.gov/DocketPDF/25/25-1018/396860/20260220123331355_PhRMA-OR–Petition%2002-20%20rtf.pdf. The petition challenges Oregon’s requirement that drug manufacturers publicly justify their pricing in reports intended to address “information asymmetries” and protect smaller retailers or consumers. It argues that compelled disclosure of pricing information violates the First Amendment and the Takings Clause by forcing the release of trade secrets. S. 8563 similarly targets “terms of sale” and “pricing differentials,” potentially requiring suppliers to disclose pricing terms offered to retailers. The Supreme Court’s resolution of whether such disclosure mandates trigger strict or intermediate scrutiny could therefore affect New York’s authority to enforce S. 8563’s reporting requirements.

[11] H.F. 2149, 94th Leg., Reg. Sess. (Minn. 2025), https://www.revisor.mn.gov/bills/94/2025/0/HF/2149.

[12] S.F. 2556, 94th Leg., Reg. Sess. (Minn. 2025), https://www.revisor.mn.gov/bills/94/2025/0/SF/2556.

[13] Lisa Kaczke, Supporters Say Pricing Fairness Bill Would Level Playing Field for Local Grocers, Minn. House of Reps.: Session Daily (Mar. 26, 2025), https://www.house.mn.gov/sessiondaily/Story/18654.

[14] S.B. 1226, 57th Leg., 2d Reg. Sess. (Ariz. 2026), https://www.azleg.gov/legtext/57leg/2R/bills/SB1226P.pdf; H.B. 2948, 57th Leg., 2d Reg. Sess. (Ariz. 2026), https://www.azleg.gov/legtext/57leg/2R/bills/HB2948P.pdf.

[15] H.R. 7514, 2026 Gen. Assemb., Jan. Sess. (R.I. 2026), https://webserver.rilegislature.gov/BillText/BillText26/HouseText26/H7514.pdf.

[16] L.D. 1894, 132d Leg., 1st Spec. Sess. (Me. 2025) (as introduced May 5, 2025), https://legislature.maine.gov/legis/bills/getPDF.asp?paper=HP1265&item=1&snum=132.

[17] L.D. 1894, 132d Leg., 1st Spec. Sess. (Me. 2023) (Comm. Amend. H-____), https://legislature.maine.gov/legis/bills/getPDF.asp?paper=HP1265&item=2&snum=132.

[18] See, e.g., Timothy J. Muris & Jonathan E. Nuechterlein, Chicago and Its Discontents, 87 U. Chi. L. Rev. 495 (2020); see also John M. Yun, From Discount to Discrimination: The Strange Economics of Anti-Competitive Antitrust, Truth on the Mkt. (Feb. 23, 2026), https://truthonthemarket.com/2026/02/23/from-discount-to-discrimination-the-strange-economics-of-anti-competitive-antitrust.

[19] D. Daniel Sokol, Analyzing Robinson-Patman, 83 Geo. Wash. L. Rev. 2064 (2015).

[20] Antitrust Modernization Comm’n, Report and Recommendations 317–18 (2007), https://govinfo.library.unt.edu/amc/report_recommendation/amc_final_report.pdf.

[21] Brooke Grp. Ltd. v. Brown & Williamson Tobacco Corp., 509 U.S. 209, 220 (1993).

[22] Volvo Trucks N. Am., Inc. v. Reeder-Simco GMC, Inc., 546 U.S. 164 (2006).

[23] FTC v. PepsiCo, Inc., FTC Docket No. 9429 (2025).

[24] Press Release, Fed. Trade Comm’n, FTC Dismisses Lawsuit Against PepsiCo (May 22, 2025), https://www.ftc.gov/news-events/news/press-releases/2025/05/ftc-dismisses-lawsuit-against-pepsico.

[25] Complaint, FTC v. S. Glazer’s Wine & Spirits, LLC, No. 8:24-cv-02684 (C.D. Cal. Dec. 12, 2024).

[26] Id. (order denying defendant’s motion to dismiss) (Apr. 17, 2025).

[27] LA Int’l Corp. v. Prestige Brands Holdings, Inc., Nos. 24-3776, 24-5009, 24-5227, slip op. (9th Cir. Feb. 24, 2026), https://cdn.ca9.uscourts.gov/datastore/opinions/2026/02/24/24-5009.pdf.

[28] See, e.g., Bruce’s Juices, Inc. v. Am. Can Co., 330 U.S. 743, 752–53 (1947) (“The [Robinson-Patman Act] does not prohibit all quantity discounts, but expressly permits them under certain conditions. … Quantity discounts are among the oldest, most widely employed, and best-known discount practices.”); see also Statement of Comm’r Alvaro M. Bedoya, Joined by Chair Lina M. Khan & Comm’r Rebecca Kelly Slaughter, In the Matter of S. Glazer’s Wine & Spirits, LLC, FTC File No. 211-0155 (Dec. 12, 2024), https://www.ftc.gov/system/files/ftc_gov/pdf/statement-bedoya-joined-by-khan-slaughter-southern-glazers.pdf (“Section 2(a) clarified that the Clayton Act’s prohibition on price discrimination permits price differences reflecting bona fide efficiencies from bulk purchases—specifically, lower costs associated with the ‘manufacture, sale, or delivery’ of large quantities of goods.”).

[29] Texaco Inc. v. Hasbrouck, 496 U.S. 543 (1990).

[30] Sponsor’s Memo, supra note 2.

[31] For a history of the development of the economics of the “waterbed effect,” see Eric Fruits, Sloshing Around with the ‘Waterbed Effect’, Truth on the Mkt. (Sept. 5, 2023), https://truthonthemarket.com/2023/09/05/sloshing-around-with-the-waterbed-effect.

[32] Roman Inderst & Tommaso M. Valletti, Buyer Power and the “Waterbed Effect”, 59 J. Indus. Econ. 1 (2011).

[33] DeHoog v. Anheuser-Busch InBev, SA/NV, No. 1:15-CV-02250-CL, 2016 U.S. Dist. LEXIS 137759 (D. Or. July 22, 2016).

[34] U.K. Competition Comm’n, Safeway plc and Asda Group Ltd. (owned by Wal-Mart Stores Inc); Wm Morrison Supermarkets plc; J Sainsbury plc; and Tesco plc: A Report on the Mergers in Contemplation ¶ 2.246 (2003), https://webarchive.nationalarchives.gov.uk/ukgwa/20120119144540/http://www.competition-commission.org.uk/rep_pub/reports/2003/481safeway.htm#full.

[35] U.K. Office of Fair Trading, The Grocery Market: The OFT’s Reasons for Making a Reference to the Competition Commission (OFT845) ¶ 6.13 (2006), https://assets.publishing.service.gov.uk/media/555de47840f0b669c4000141/oft845.pdf.

[36] Bruce H. Kobayashi & Timothy J. Muris, Stop Making Sense: Reviving the Robinson-Patman Act and the Economics of Intermediate Price Discrimination, Competitive Enter. Inst. (Feb. 18, 2026), https://cei.org/wp-content/uploads/2026/02/Stop-Making-Sense-260218-FINAL.pdf.

[37] See, e.g., Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini & Florian Tramèr, Large-scale Online Deanonymization with LLMs, arXiv (Feb. 18, 2026), https://doi.org/10.48550/arXiv.2602.16800.

[38] George J. Stigler, A Theory of Oligopoly, 72 J. Pol. Econ. 44 (1964).

[39] Organisation for Econ. Co-operation & Dev. (OECD), Information Exchanges Between Competitors under Competition Law, DAF/COMP(2010)37, at 10 (2010), https://www.oecd.org/content/dam/oecd/en/publications/reports/2011/07/information-exchanges-between-competitors-under-competition-law_bd644d8b/327f7dd3-en.pdf (“Notwithstanding its benefits, enhanced transparency can also facilitate the attainment of collusive equilibria among competitors or result in non-coordinated anticompetitive effects.”).

[40] Note by the United States, Roundtable on Information Exchanges Between Competitors under Competition Law, DAF/COMP/WD(2010)117 ¶ 3 (Oct. 2010), https://www.ftc.gov/sites/default/files/attachments/us-submissions-oecd-and-other-international-competition-fora/1010informationexchanges.pdf (“The antitrust concern is that information exchanges may facilitate anticompetitive harm by improving competing sellers’ ability either to collude or to tacitly coordinate in ways that lessen competition. For example, exchanges about price may lead to illegal price coordination. Information exchanges may also facilitate collective behavior by downstream firms against upstream suppliers.”).

[41] Zoë Cullen & Bobak Pakzad-Hurson, Equilibrium Effects of Pay Transparency, 91 Econometrica 765 (2023) (“Our model predicts that transparency reduces workers’ individual bargaining power, leading to lower average wages. Employers credibly refuse to offer higher wages to any one worker to avoid triggering costly renegotiations with others.”).

[42] Antitrust Modernization Comm’n, supra note 20, at 322.

[43] Emek Basker, The Causes and Consequences of Wal-Mart’s Growth, 21 J. Econ. Persp. 177 (2007).

[44] Id. at 188.

[45] Jerry Hausman & Ephraim Leibtag, Consumer Benefits from Increased Competition in Shopping Outlets: Measuring the Effect of Wal-Mart, 22 J. Appl. Econ. 1157 (2007).

[46] Yun, supra note 18.

[47] Michele Ver Ploeg et al., Access to Affordable and Nutritious Food: Measuring and Understanding Food Deserts and Their Consequences, USDA Econ. Research Serv., Report to Congress (AP-036, 2009), https://ers.usda.gov/sites/default/files/_laserfiche/publications/42711/12716_ap036_1_.pdf?v=32613.

[48] Id.

[49] Hunt Allcott, Rebecca Diamond, Jean-Pierre Dubé, Jessie Handbury, Ilya Rahkovsky & Molly Schnell, Food Deserts and the Causes of Nutritional Inequality, 134 Q.J. Econ. 1793 (2019).

[50] Steven Cummins, Ellen Flint & Stephen A. Matthews, New Neighborhood Grocery Store Increased Awareness of Food Access but Did Not Alter Dietary Habits or Obesity, 33 Health Aff. 283 (2014).

[51] Hausman & Leibtag, supra note 44.

Executive Summary

Liberal democracies rest on five interlocking pillars: separation of powers, subsidiarity, the rule of law, electoral accountability, and freedom of speech. Developed over centuries—from Montesquieu and James Madison to Charles Tiebout and Elinor Ostrom—these institutions reflect a common insight: liberty depends not on the virtue of rulers but on institutional design. Authority must be fragmented, decisions kept close to those they affect, and citizens given the tools to contest and correct those who govern.

This paper identifies a structural threat to that architecture: “Democratically Deficient Organizations,” or “DoDOs.” A DoDO is not a single institution but an ecosystem of intergovernmental bodies, nongovernmental organizations (NGOs), foundations, and corporations that collectively shape policy while remaining largely insulated from democratic accountability. These networks share four defining features: they exercise meaningful policy-shaping authority, lack direct electoral accountability, operate through opaque decision-making processes, and expand their mandates and budgets in response to the problems they claim to address.

The DoDO dynamic has emerged through a gradual upward drift of decision-making authority. As power moves from local to national and from national to supranational institutions, the distance between citizens and decision makers grows and accountability weakens. A voter can attend a town council meeting or vote out a mayor; no “citizen of the world” can meaningfully hold the World Health Organization (WHO) Executive Board or the Organisation for Economic Co-operation and Development (OECD) Committee on Fiscal Affairs to account. As William Niskanen’s analysis suggests, institutions facing weak oversight tend naturally to expand their budgets and jurisdictions.

Several policy domains illustrate how this system operates in practice. Development aid often weakens the incentives of recipient governments to pursue institutional reform, while a substantial share of aid spending flows back to contractors and NGOs in donor countries. Environmental treaties—such as the Convention on International Trade in Endangered Species (CITES) and the Cartagena Protocol on Biosafety—can undermine conservation incentives and technological innovation, while sustaining large advocacy and compliance ecosystems. Global health governance has similarly expanded beyond its original focus on infectious disease, with initiatives such as the Framework Convention on Tobacco Control opposing harm-reduction products that independent scientific assessments find far less harmful than combustible tobacco.

The global drug-prohibition regime demonstrates that the DoDO dynamic transcends ideology. International treaties and bureaucratic institutions have entrenched prohibition despite significant social costs, including mass incarceration and widespread violence. Efforts to disrupt drug-related financial flows have produced expansive anti–money laundering regimes that impose vast compliance costs with limited evidence of effectiveness. In fiscal policy, the OECD’s campaign against “harmful tax competition”—including the Base Erosion and Profit Shifting (BEPS) initiative and the Common Reporting Standard—illustrates how international coordination can suppress the jurisdictional competition that disciplines governments.

Two theoretical frameworks help explain why such policies persist. Timur Kuran and Cass Sunstein’s concept of availability cascades describes how alarming claims can generate self-reinforcing policy cycles: NGOs highlight risks, media amplify them, politicians respond with new mandates, international treaties lock in those responses, and expanding bureaucracies reinforce the original claims. Bruce Yandle’s “Bootleggers and Baptists” model shows how moral advocates and economic beneficiaries form durable coalitions that sustain regulation. Once policy debates become moralized—saving the planet, protecting children, fighting poverty—dissent becomes stigmatized, rather than debated.

The brief concludes with an analogy from parasitology. The lancet liver fluke (Dicrocoelium dendriticum) does not kill its ant host; it alters a single behavior to ensure its own survival. Likewise, the DoDO complex does not abolish democratic institutions but subtly redirects them, weakening the mechanisms through which citizens correct failing policies.

The remedy is not to dismantle the organizations involved but to interrupt the self-reinforcing logic that sustains them. Reforms should restore subsidiarity by returning authority to the level closest to those affected, impose sunset clauses and periodic reauthorization on international commitments, end government funding of NGOs that lobby for the very government policies that sustain them, increase transparency in treaty negotiations, and re-anchor international cooperation in jurisdictional competition and local experimentation. Subsequent papers in this series will examine these reforms across specific policy domains.

I. Introduction: Meet the DoDOs

“Giving money and power to government is like giving whiskey and car keys to teenage boys.”

— P. J. O’Rourke

Those of us privileged to live in liberal democracies often take the system for granted. That complacency obscures a basic reality: the institutions that sustain democratic governance are fragile. Each pillar can erode, often gradually and quietly, and once institutional corrosion spreads, rebuilding becomes difficult. As Lord Acton warned Bishop Mandell Creighton in 1887, “power tends to corrupt, and absolute power corrupts absolutely.”^[1]

The danger is particularly acute where liberal democracy lacks deep institutional roots. Venezuela offers a stark illustration. Within a single generation, Hugo Chávez and his successors weakened or dismantled the core safeguards of democratic governance. The judiciary fell under political control, the legislature was sidelined, the press faced systematic pressure, and elections increasingly served as a ritual of legitimacy, rather than a meaningful mechanism of political choice.^[2]

This issue brief—the first in a series—argues that liberal democracies face a subtler but still serious threat from what I call “Democratically Deficient Organizations,” or “DoDOs.” These are coalitions of intergovernmental bodies, non-governmental organizations (NGOs), foundations, and corporations that together exercise significant policy-shaping authority, while remaining largely insulated from democratic accountability.

The concept is not simply a label for institutions or policies I dislike. Rather, it describes a specific structural phenomenon: an ecosystem of organizations that collectively meet four criteria.

First, they exercise meaningful policy-shaping authority. This authority may arise through binding treaties, regulatory standards, funding conditions, or the creation of influential technical norms.

Second, they lack direct democratic accountability. No electorate can remove their leadership or reverse their decisions through the ballot box.

Third, their decisionmaking processes remain largely insulated from external scrutiny and contestation. Proceedings are closed, dissenting views receive little institutional space, and affected populations often lack a meaningful voice.

Fourth, they exhibit self-reinforcing institutional incentives. Budgets, mandates, and influence tend to expand in response to the very problems these organizations are charged with solving, creating a structural bias toward growth, rather than resolution.

The phenomenon is not merely theoretical. Across domains as varied as development aid, environmental regulation, public-health governance, and international taxation, the DoDO ecosystem has produced policies that persist long after evidence of failure accumulates.

Not every international organization qualifies as a DoDO. The North Atlantic Treaty Organization (NATO), for example, operates under a clearly defined mandate subordinate to the democratic decisions of member states. Nor does every nongovernmental organization fit the model. A local charity that builds wells in a particular village and answers to identifiable donors operates within a recognizable chain of accountability.

The problem arises when organizations that fail all four criteria operate in concert. Intergovernmental bodies, NGOs, foundations, and commercial actors can form mutually reinforcing networks that collectively wield authority no single institution could claim on its own. The pathology lies not in any individual organization, but in the ecosystem they create together.^[3]

I first developed this concept roughly 20 years ago while preparing brief introductory remarks for a fundraising dinner hosted by the International Policy Network. The keynote speaker was the late political humorist P. J. O’Rourke. I wanted to explain why the organization existed. My observation was that policymaking increasingly occurred through coalitions of institutions, each with only a weak claim to democratic accountability, but which together created a powerful appearance of legitimacy. I called them Democratically Deficient Organizations, or DoDOs.

After the event, P. J. told me he liked the concept and urged me to write it up. I promised that I would. This paper series is a somewhat belated fulfillment of that promise.[4]

The purpose of this first brief is to establish the conceptual foundation for the broader project. Section II provides an intellectual history of the mechanisms developed to constrain delegated power—from Montesquieu through James Madison and Alexis de Tocqueville to the economic analysis of Charles Tiebout and Elinor Ostrom. Section III traces how authority has gradually shifted upward over the past century from local institutions to supranational governance, illustrating each stage with concrete examples. Section IV then draws these strands together through an analogy from parasitology, identifying the structural features that make the DoDO complex self-sustaining and outlining the reforms explored in subsequent papers in this series.

II. The Architecture of Democratic Accountability

“Sed quis custodiet ipsos custodes?”

— Juvenal, Satires VI^[5]

Juvenal’s famous question (“But who will guard the guardians?”) captures the enduring problem of delegated power: how to ensure that those who act on our behalf do not come to act primarily for themselves. Political thinkers have wrestled with this question for more than two millennia. No single solution exists. Instead, liberal democracies have developed a set of complementary institutional mechanisms that fragment authority, constrain discretion, and enable continuous contestation of those who govern.

Five mechanisms stand out: separation of powers, subsidiarity, the rule of law, electoral accountability, and freedom of speech. Together, they form the institutional architecture of democratic accountability.^[6]

A. Separation of Powers

The idea that liberty requires distributing government functions among distinct institutions received its most influential formulation from Charles-Louis de Secondat, Baron de Montesquieu, in “De l’Esprit des Lois” (1748). In Book XI, Chapter 6, Montesquieu argued that liberty disappears when legislative, executive, and judicial power concentrate in the same hands:^[7]

When the legislative and executive powers are united in the same person, or in the same body of magistrates, there can be no liberty; because apprehensions may arise, lest the same monarch or senate should enact tyrannical laws, to execute them in a tyrannical manner.

Montesquieu derived this insight from his observations of the English constitution during his travels in 1729–1731. His interpretation was, by scholarly consensus, somewhat idealized. Britain never practiced a strict separation of powers. Yet that very idealization made his formulation more radical and more widely applicable than the British system itself.^[8]

The authors of “The Federalist Papers” adopted and refined Montesquieu’s framework. In “Federalist No. 47,” James Madison responded to critics who argued that the proposed U.S. Constitution concentrated too much authority. Madison examined Montesquieu’s argument and the British model from which it drew, concluding that separation of powers did not require complete institutional isolation. Rather, it required that no branch possess the whole power of another:^[9]

The accumulation of all powers legislative, executive and judiciary in the same hands, whether of one, a few or many, and whether hereditary, self appointed, or elective, may justly be pronounced the very definition of tyranny.

In “Federalist No. 51,” Madison described how the system would function in practice. Separation would not rely on “parchment barriers,” but on institutional incentives that align ambition with constitutional structure:^[10]

Ambition must be made to counteract ambition. The interest of the man must be connected with the constitutional rights of the place. It may be a reflection on human nature, that such devices should be necessary to controul the abuses of government. But what is government itself, but the greatest of all reflections on human nature? If men were angels, no government would be necessary. If angels were to govern men, neither external nor internal controuls on government would be necessary.

Madison’s key insight was institutional, rather than moral. Because people are not angels, the goal is not to find virtuous rulers, but to design institutions that channel self-interest toward mutual constraint.

The U.S. Constitution implements this insight through the allocation of authority among Congress, the president, and the federal judiciary.^[11] Variations on the same logic appear across liberal democracies—from the Westminster system’s division among Crown, Parliament, and courts to France’s séparation des pouvoirs under the Fifth Republic.

The American system adds an additional dimension: the combination of horizontal separation among branches with vertical separation among levels of government. Madison described this structure in “Federalist No. 39” as a “compound republic”—“in strictness, neither a national nor a federal constitution; but a composition of both.”^[12]

B. Subsidiarity: Keeping Power Close to the People

The vertical dimension of institutional design deserves particular attention, because it is along this axis that democratic accountability most often erodes.

Madison’s defense of federalism rested partly on scale. In “Federalist No. 10,” he argued that a large republic could control the mischief of faction because no single faction could easily dominate across such a broad polity.^[13] But this was an argument for federalism, not centralization.

As Madison emphasized in “Federalist No. 45,” federal authority would remain limited: the powers delegated to the national government were “few and defined,” while those retained by the states were “numerous and indefinite.”^[14] The Tenth Amendment enshrined this principle: “The powers not delegated to the United States by the Constitution, nor prohibited by it to the States, are reserved to the States respectively, or to the people.”^[15]

Two centuries later, economists formalized the logic behind this vertical allocation of authority. Charles Tiebout showed that when citizens can “vote with their feet”—moving to jurisdictions that better match their preferences—governments face competitive pressure to provide policies citizens value.^[16]

Elinor Ostrom extended this insight through extensive empirical research. She demonstrated that communities often manage shared resources—such as fisheries, forests, and irrigation systems—more effectively than distant central authorities, provided those communities retain the autonomy to develop rules adapted to local conditions.^[17]

Both Tiebout and Ostrom highlight a common principle: decentralization is not merely a political preference. It is an informational necessity. Local actors possess knowledge about local conditions that centralized authorities cannot replicate.^[18]

A parallel idea emerged independently in Catholic social teaching under the doctrine of subsidiarity. First articulated implicitly in Pope Leo XIII’s “Rerum Novarum” (1891) and later formulated explicitly in Pope Pius XI’s “Quadragesimo Anno” (1931), subsidiarity holds that decisions should occur at the lowest level capable of addressing them effectively:^[19]

Just as it is gravely wrong to take from individuals what they can accomplish by their own initiative and industry and give it to the community, so also it is an injustice and at the same time a grave evil and disturbance of right order to assign to a greater and higher association what lesser and subordinate organizations can do.

The European Union later nominally incorporated this principle into Article 5(3) of the Treaty on European Union (Maastricht Treaty, 1992), which states that, in areas outside its exclusive competence, the EU should act only when objectives cannot be sufficiently achieved by member states.^[20]

In practice, the principle has proved difficult to enforce. As Giandomenico Majone has shown, European institutions frequently expanded their authority beyond the bounds subsidiarity would suggest. Ambiguous treaty language and institutional incentives facilitated what Majone describes as “integration by stealth.”^[21] Indeed, as early as 1979, long before the 1992 Maastricht Treaty that officially established the European Union, David Marquand had identified the emerging “democratic deficit” in European governance—a phrase that has since entered common usage.^[22]

C. The Rule of Law

The rule of law requires that political authority operate through general, public, prospective, and stable rules, rather than through arbitrary discretion. Both rulers and citizens must remain subject to the same legal framework.

The concept traces back to antiquity. In “Politics,” Aristotle distinguished rule by law from rule by individuals, arguing that “he who bids the law rule may be deemed to bid God and Reason alone rule, but he who bids man rule adds an element of the beast.”^[23]

The Roman Republic embodied important elements of this principle. Written statutes (leges) governed public life, and magistrates operated within a framework shaped by legal scholars (jurisprudentes).[24] Unlike common-law systems, Roman law did not bind courts to precedent, which sometimes made it easier for elites to manipulate legal outcomes.[25] Over time—particularly after the fall of the republic—the rule of man gradually replaced the rule of law.[26]

In medieval Europe, the rule of law reemerged through jurisdictional competition between church and state, and through the development of merchant courts such as the pied poudre courts associated with the Champagne fairs.[27] These evolving constraints on sovereign authority found their most famous expression in Magna Carta (1215), which established that even the king was subject to “the law of the land.”^[28]

John Locke later provided the modern theoretical foundation for the rule of law in his “Second Treatise of Government” (1689):^[29]

Absolute arbitrary power, or governing without settled standing laws, can neither of them consist with the ends of society and government … For the end of government is the good of mankind; and which is best attained by standing laws, to which every one may know his duty, and be safe and secure within the limits of the law; and not by extemporary dictates and undetermined resolutions of men.

Friedrich A. Hayek deepened this analysis in “The Road to Serfdom” (1944). Hayek argued that centralized economic planning conflicts with the rule of law because governments that allocate resources according to predetermined outcomes must exercise discretionary authority. That discretion undermines the generality and neutrality that legal rules require.^[30]

In “The Constitution of Liberty” (1960), Hayek extended the argument further, contending that the rule of law functions not merely as a procedural requirement, but as a constraint on legislative substance. Laws must be general, certain, and equally applied. Otherwise, they become commands issued under the guise of law.^[31]

D. Electoral Accountability

Economists frame the challenge of democratic governance as a principal-agent problem. Citizens (the principals) delegate authority to elected officials (the agents). Yet agents may pursue their own interests, rather than those of the principals who empower them.^[32]

Elections provide the most direct mechanism for aligning those interests. Within the principal-agent framework, elections perform two key functions. First, they serve as a selection mechanism. Voters choose candidates whose preferences and competence they believe best serve the public interest. Second, elections create discipline. The possibility of removal from office encourages incumbents to remain responsive to voter preferences.^[33]

Electoral accountability remains imperfect. Voters face substantial information asymmetries. They observe outcomes such as prosperity, security, or public health, but cannot easily distinguish the effects of policy from external shocks. Rational ignorance compounds the problem: because any single vote rarely determines an election, the incentive for individuals to acquire detailed political knowledge remains limited.^[34]

Even so, competitive elections exert meaningful constraints on elected officials. Empirical evidence consistently associates regular elections with improved governance outcomes.^[35]

The strength of electoral accountability also depends on proximity. A mayor faces reelection before voters who directly experience the consequences of local policy. A national legislator answers to a more diffuse electorate. Bureaucrats—career administrators within government agencies—do not stand for election at all. Officials within international organizations operate even further removed from direct democratic control.

This widening distance between decisionmaker and voter weakens accountability—a dynamic that becomes central to the DoDO phenomenon examined in later sections.

E. Freedom of Speech

None of these institutional safeguards can function without freedom of expression. Citizens cannot hold their representatives accountable without the ability to exchange information, criticize policy, and organize political opposition.

Alexis de Tocqueville identified freedom of the press as indispensable to republican government. In “Democracy in America,” he argued that press freedom allowed citizens to coordinate resistance to abuses of power that might otherwise remain invisible:^[36]

In countries where the doctrine of the sovereignty of the people ostensibly prevails, the censorship of the press is not only dangerous, but absurd.

John Stuart Mill later advanced the epistemic defense of free speech in “On Liberty” (1859). Mill contended that suppressing opinions harms society, whether those opinions prove true or false. Only through the collision of competing ideas, Mill argued, can truth emerge and remain intellectually alive.^[37]

Justice Louis Brandeis expressed the same principle in his famous concurrence in Whitney v. California:^[38]

Those who won our independence believed … that public discussion is a political duty; and that this should be a fundamental principle of the American government. They recognized the risks to which all human institutions are subject. But they knew that order cannot be secured merely through fear of punishment for its infraction; that it is hazardous to discourage thought, hope and imagination; that fear breeds repression; that repression breeds hate; that hate menaces stable government.

These insights illuminate the broader argument of this paper. Democratic institutions rely not only on formal structures of accountability, but also on the open exchange of ideas that allows those structures to function. When dissent becomes stigmatized or policy debate becomes moralized, the epistemic foundations of democratic self-correction weaken.

As the following sections will argue, the DoDO ecosystem often produces precisely these effects.

III. The Drift of Power

“The whole aim of practical politics is to keep the populace alarmed … by menacing it with an endless series of hobgoblins, most of them imaginary.”

— H. L. Mencken, In Defense of Women (1918)

Democracy means rule by the people. In theory, democratic authority flows in a clear direction. Citizens authorize governments. Governments delegate tasks to agencies. Agencies act on behalf of the public, and when they act badly, their authority can be withdrawn, constrained, or redirected. But this system depends on friction: elections that matter, oversight that bites, and consequences that are felt. When those frictions weaken, authority does not collapse. It drifts.^[39]

Over the course of the past century, power has gradually accumulated at higher levels of aggregation than would be optimal from a principal-agent perspective. At each step upward, the distance between principal and agent grows, mechanisms of accountability weaken, and opportunities for self-serving behavior multiply.

A voter in a small town can attend a council meeting, speak to the mayor, or run for office. A citizen of a nation state has the ballot box, judicial review, and constitutional protections. But a “citizen of the world” has no meaningful mechanism by which to hold the World Health Organization (WHO) Executive Board, the Organisation for Economic Co-operation and Development’s (OECD) Committee on Fiscal Affairs, or the World Bank’s Board of Governors to account. These bodies answer, if at all, to the executives of member states, who themselves are often only loosely accountable to their citizens on matters of international policy.^[40]

This upward drift has not occurred through a single act of usurpation. It has been gradual, cumulative, and often well intentioned. Each delegation of authority is typically justified by plausible arguments: economies of scale, coordination needs, or cross-border externalities. Yet the cumulative effect is a system in which decision making has steadily moved away from those most affected by it.

Failure rarely produces correction. Instead, it generates demands for further centralization. As William Niskanen demonstrated, bureaucracies are not neutral servants of the public interest. They respond to incentives, and the dominant incentive within bureaucratic systems is to expand budgets, jurisdiction, and influence.^[41]

At the same time, non-representative actors—including NGOs, philanthropic foundations, and corporations—have sought to influence policy at every level of government. Their influence is often greatest at the supranational level, where democratic representation is weakest and concentrated interests can most easily override diffuse public interests.^[42]

The result is the DoDO complex: democratically deficient governments and supranational bodies operating in concert with inherently non-democratic NGOs, foundations, and businesses to promote agendas (and, ultimately, legislation and funding) that serve the interests of the DoDO ecosystem, rather than the people they claim to represent.

The following examples illustrate how this dynamic now operates across multiple policy domains.

A. Aid Without Accountability

Development aid is often justified as a tool for promoting economic growth and poverty reduction. In practice, it has frequently produced the opposite effect.

By providing large and relatively unconditional flows of external funding, aid reduces the incentive for recipient governments to pursue institutional reforms that historically underpin sustained development—secure property rights, independent courts, trade liberalization, and accountable fiscal systems. When governments finance substantial portions of their budgets through foreign assistance rather than domestic taxation, political pressure to improve governance weakens.

Scholars have long noted that this dynamic can undermine the social contract between governments and citizens. Political leaders become accountable primarily to foreign donors, rather than domestic taxpayers.[43] Empirical research likewise finds that large aid inflows often correlate with weaker institutional development and slower improvements in governance.[44]

To be clear, aid does sometimes support growth. But evidence suggests that aid tends to produce positive long-term effects only in countries that already maintain sound fiscal, monetary, and trade policies.^[45]

At the same time, much aid spending never reaches recipient economies. Significant portions fund contractors, consultants, and implementing organizations in donor countries. In the United States, the majority of funding administered by the U.S. Agency for International Development (USAID) historically flowed back to American entities through procurement contracts, grants, and technical-assistance programs. Analyses of USAID procurement data indicate that roughly 60–70% of USAID funding was awarded to U.S.-based firms, universities, and nonprofit organizations, while only a small share went directly to local organizations in recipient countries.[46]

This structure has helped create a large DoDO aid industry composed of development contractors, consulting firms, NGOs, and research institutions whose revenues depend on continued aid spending.[47] Unsurprisingly, these organizations rank among the most vocal advocates of foreign-aid programs.[48]

B. Conservation Without Incentives

A similar pathology appears in many international environmental agreements.

Treaties such as the Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES) and the Cartagena Protocol on Biosafety were promoted as mechanisms to protect biodiversity and prevent environmental harm. Yet their design has often weakened incentives for conservation and technological innovation.

Environmental-advocacy organizations, research institutes, and policy consultants play a central role in promoting such agreements. By highlighting environmental risks and mobilizing public opinion, these groups generate the political momentum needed for governments to negotiate treaties and create regulatory regimes.[49] Yet the policy tools these agreements rely on—primarily trade restrictions and precautionary regulatory frameworks—often undermine the economic incentives that make conservation effective.

For example, CITES attempts to protect endangered species by restricting international trade in wildlife products. While well intentioned, these restrictions reduce the in situ economic value of wildlife—its value while alive in its natural habitat. This reduces incentives for landowners and local communities to invest in habitat protection or tolerate wildlife that imposes real costs.[50]

Similarly, the Biosafety Protocol imposes extensive regulatory requirements on genetically modified crops, despite broad scientific consensus that such crops can improve yields, reduce pesticide use, and relieve pressure on natural ecosystems.[51]

As with development aid, these environmental regulatory regimes sustain a substantial ecosystem of advocacy organizations, consultants, compliance specialists, and implementing agencies whose revenues depend on the expansion and enforcement of international environmental governance. Unsurprisingly, these groups rank among the most enthusiastic proponents of new treaties and regulatory initiatives—even when the policies they promote undermine the very conservation outcomes they claim to advance.[52]

C. The Global Health Governance Complex

International public-health governance has followed a similar trajectory. Institutions such as the World Health Organization (WHO) were originally established to coordinate international responses to infectious diseases and to share scientific information among national health authorities. Over time, however, the organization’s mandate has expanded substantially, and global health policy increasingly emerges from networks of international agencies, philanthropic foundations, NGOs, and private contractors. Advocacy organizations and research groups often help shape this agenda by highlighting particular health risks and mobilizing public support for regulatory responses.[53]

The WHO’s Framework Convention on Tobacco Control (FCTC) illustrates the DoDO dynamic with particular clarity. Bloomberg Philanthropies is one of the largest private funders of global tobacco-control initiatives and provides substantial support for programs promoting implementation of the FCTC.[54] Through the Bloomberg Initiative to Reduce Tobacco Use, the foundation finances a network of advocacy organizations, research institutes, and implementing partners, many of which participate in the Global Alliance for Tobacco Control (formerly the Framework Convention Alliance). These NGOs play a central role in producing reports, lobbying governments, training policymakers, and coordinating advocacy during FCTC negotiations.

Much of the advocacy surrounding the FCTC has focused on strict regulation of alternative nicotine products, including snus and e-cigarettes. WHO policy documents and reports presented to the Conference of the Parties have repeatedly recommended measures such as advertising bans, restrictions on use, and, in some cases, prohibition of e-cigarettes.[55] Yet independent scientific assessments by bodies such as the Royal College of Physicians and Public Health England have found such alternatives far less harmful than combustible tobacco.^[56] Restrictions on alternative nicotine products inevitably discourage the adoption of harm-reduction technologies that could reduce smoking-related disease.[57]

More broadly, the expansion of global health governance has helped sustain a large ecosystem of advocacy organizations, research institutes, consulting firms, and implementing partners whose funding depends on the continued growth of international health programs.[58] The result is a familiar pattern: mandate creep from communicable diseases to non-communicable diseases, and from tobacco to obesity, alcohol, and sugar. Each expansion, justified by moral urgency, reduces the scope for national democratic decision making.

D. The Global Prohibition Complex

The global drug-prohibition regime demonstrates that the DoDO problem transcends ideology. Here is a fundamentally conservative policy project, rooted in moral absolutism and sustained by law enforcement, yet it operates through precisely the mechanisms that characterize many left-leaning DoDOs: unaccountable bureaucracies, international coordination, expert commissions, and the suppression of democratic alternatives.

The global prohibition system, formalized in the 1961 U.N. Single Convention on Narcotic Drugs and expanded through subsequent treaties, has produced profound institutional and human costs.^[59] The U.S. Drug Enforcement Administration (DEA) has grown from a few hundred agents in 1973 into a multibillion-dollar bureaucracy with offices in dozens of countries. At the international level, the U.N. Office on Drugs and Crime (UNODC) and the International Narcotics Control Board provide legitimacy and coordination but operate without meaningful democratic mandates. Countries that seek to liberalize drug policy risk violating treaty obligations, which effectively constrains democratic experimentation.

The human consequences have been severe. More than 2 million people are incarcerated in the United States, while cartel violence has killed more than 460,000 people in Mexico since 2006. Meanwhile, prohibition has normalized expansive surveillance tools—including civil asset forfeiture, no-knock warrants, and extensive financial monitoring.^[60]

Efforts to disrupt drug-related financial flows have further expanded this institutional apparatus. Governments introduced anti–money laundering (AML) regimes—later broadened to include countering the financing of terrorism and proliferation financing—that impose hundreds of billions of dollars in compliance costs and affect hundreds of millions of people worldwide.[61] Yet despite these costs, evidence that such measures substantially reduce drug trafficking or terrorist finance remains limited.^[62]

E. The War on Tax Competition

Since the mid-1990s, the Organisation for Economic Co-operation and Development (OECD) has led an international campaign against what it calls “harmful tax competition.”^[63] In practice, the campaign seeks primarily to protect high-tax jurisdictions from competitive pressure on corporate income taxes—ironically the very form of taxation the OECD itself identifies as most harmful for economic growth.[64]

Meanwhile, competition among jurisdictions to attract mobile capital has placed sustained downward pressure on statutory corporate tax rates across advanced economies. Despite these declining rates, total tax revenues as a share of GDP in many countries have remained stable or increased.[65] In this sense, what the OECD describes as harmful tax competition often functions as beneficial fiscal discipline.

The centerpiece of the OECD campaign has been the Base Erosion and Profit Shifting (BEPS) initiative, which has produced increasingly complex rules designed to limit companies’ ability to allocate profits across jurisdictions. In parallel, the OECD established the Common Reporting Standard (CRS), which requires automatic exchange of financial-account information among tax authorities. These rules have eroded financial privacy, while imposing substantial compliance costs on financial institutions and taxpayers.^[66]

The Financial Action Task Force (FATF), originally established by the G7 in 1989 to combat drug-related money laundering, has become an important enforcement mechanism within this broader agenda. Through its “grey list,” FATF pressures jurisdictions—particularly smaller or lower-tax countries—to adopt increasingly expansive transparency and regulatory standards.[67]

These initiatives operate within a wider ecosystem that includes advocacy NGOs, research institutes, and international bureaucracies. Organizations such as the Tax Justice Network (TJN), Oxfam, and Christian Aid have played prominent roles in promoting stricter international tax coordination and in opposing tax competition. Through reports, media campaigns, and policy advocacy, these groups have pressed governments to adopt measures such as automatic information exchange, global minimum taxes, and expanded anti-avoidance rules.[68] Coordinating networks such as the Global Alliance for Tax Justice and the Financial Transparency Coalition amplify these efforts.[69]

Public institutions themselves increasingly fund research organizations that advocate similar policies. For example, the European Commission’s Directorate-General for Taxation and Customs Union (DG TAXUD) supports the EU Tax Observatory, a research institute whose reports frequently advocate stronger international tax coordination and criticize low-tax jurisdictions.

Taken together, these institutions form a policy network that promotes rules designed to limit fiscal competition among states. Ironically, such competition is precisely the mechanism economists like Charles Tiebout identified as a discipline on government—allowing taxpayers and firms to “vote with their feet” by moving to jurisdictions that better match their preferences.

F. The Drift Toward Democratic Deficit

A note on causation is warranted. The DoDO dynamic does not explain every policy failure discussed above.

Development aid, for example, emerged partly from the “financing gap” theory, which held that poor countries lacked sufficient domestic savings to finance the investment needed for economic growth.[70] Early critics, however, warned that large financial transfers to poorly governed states would strengthen political elites and weaken incentives for institutional reform. As Peter Bauer observed, aid often amounted to “a process by which poor people in rich countries are taxed to support rich people in poor countries.”[71]

Environmental governance offers similar examples. The International Whaling Commission, created by the 1946 International Convention for the Regulation of Whaling to ensure sustainable whale harvesting, adopted a quota system based on the “Blue Whale Unit,” which treated different whale species as interchangeable according to oil yield. This system encouraged fleets to target the largest whales first, producing a pattern of serial depletion that devastated whale populations, before the system was ultimately abandoned in the 1970s.[72]

Public health governance also provides examples of policy failure unrelated to the DoDO dynamic. Following the 1978 Alma-Ata Declaration, the WHO and many governments embraced the ambitious goal of “Health for All” through universal primary health care. In many developing countries, however, the strategy proved impractical because it required administrative capacity, funding, and political commitment that local health systems did not possess.[73]

These examples illustrate several distinct sources of policy failure: structural governance problems, flawed policy design, and ideological commitments. Yet in many of the cases described earlier—and in many others explored in subsequent papers—the DoDO dynamic clearly plays a role. Self-reinforcing networks of international organizations, NGOs, foundations, and contractors can insulate failing policies from democratic correction, while creating constituencies with strong incentives to preserve them, regardless of outcomes.

These patterns are not isolated. According to Freedom House, the world has experienced 19 consecutive years of decline in political rights and civil liberties. In 2024, 60 countries deteriorated while only 34 improved, with freedom of expression declining most sharply.^[74]

The causes of democratic decline are complex. Authoritarian populism, political polarization, and economic stagnation all contribute. Yet the upward migration of decisionmaking authority to unaccountable institutions—supported by networks of NGOs, foundations, and corporations with aligned incentives—almost certainly plays a role. Each transfer of authority away from accountable institutions weakens the habits, expectations, and norms on which democratic governance ultimately depends.

IV. The DoDO Life Cycle — Why It Persists and How to Break It

“Nations stumble upon establishments which are indeed the result of human action, but not the execution of any human design.”

— Adam Ferguson, An Essay on the History of Civil Society (1767)

The DoDO ecosystem is the product of human action, but not of human design. It evolved. Evolutionarily stable strategies, however, are not always optimal for the systems in which they arise. In that respect, the DoDO complex resembles the life cycle of the lancet liver fluke (Dicrocoelium dendriticum).

The parasite begins as tiny eggs deposited in the excreta of grazing herbivores. Snails inadvertently ingest the eggs while feeding. Inside the snail, the parasite’s larvae (the cercariae) hatch and multiply. The snail then expels them in small spheres of mucus that glisten in the sunlight like drops of dew.

Ants are drawn to these globules and consume them. Most of the cercariae form cysts in the ants’ abdomens and remain relatively harmless. One, however, migrates to the ant’s brain.

There, the parasite does not kill the ant or seize full control. Instead, it alters a single behavior. As evening temperatures fall, the infected ant climbs a blade of grass and clamps down with its mandibles, remaining there through the night. When morning warmth returns, grazing animals—sheep, cattle, goats, even alpaca—consume the grass along with the immobilized ant, returning the parasite to its final host, where it matures and lays eggs.^[75]

No single host in this sequence behaves maliciously. The snail behaves as snails do. The ant behaves almost as ants do. The cow grazes as cows always have. The pathology lies not in any single organism, but in the system: a chain of hosts whose ordinary functions are subtly repurposed to serve something else.

The analogy does not imply conspiracy. The lancet liver fluke does not plot; neither, in most cases, do the actors within the DoDO complex. Aid workers believe they are helping the poor. Environmental campaigners believe they are protecting the planet. WHO officials believe they are safeguarding public health.

The point is systemic dynamics. Like the hosts in the fluke’s life cycle, the institutions that comprise the DoDO ecosystem often serve legitimate purposes when viewed individually. Viewed as a system, however, their functions become subtly redirected toward sustaining and expanding the complex itself. Just as the fluke alters the ant’s behavior without killing it, the DoDO complex does not abolish democratic institutions; it changes their behavior just enough to sustain its own survival.

If the analogy holds, the remedy is not to destroy the hosts. Snails, ants, and cattle all serve useful purposes. The task is to interrupt the life cycle—to restore the accountability mechanisms that have been co-opted and break the feedback loops that sustain the system. To do that, we must first understand how those loops operate.

A. Availability Cascades and Policy Entrenchment

Timur Kuran and Cass Sunstein describe how public perceptions of risk can become self-reinforcing through “availability cascades.”^[76] A cascade begins when an idea gains salience—through a dramatic event, a well-funded campaign, or repeated exposure. As the idea becomes more “available” in public consciousness, advocates amplify it, media coverage expands, and policymakers increasingly treat it as established truth.

This dynamic exploits well-known cognitive biases. The availability heuristic leads people to judge risks by how easily examples come to mind. Conformity effects lead individuals to align their views with what they perceive to be the prevailing consensus.

The DoDO ecosystem functions, in effect, as an institutional engine for generating such cascades. NGOs emphasize problems, often in the most alarming terms, because alarm generates donations and attention. Journalists amplify these narratives because alarm attracts readers. Politicians respond because alarm generates votes. Foundations fund research that confirms the concern, and agencies cite that research to justify expanded mandates.

Kuran and Sunstein describe those who benefit from this process as “availability entrepreneurs”—actors who profit from amplifying risk perceptions, regardless of their accuracy.^[77]

The feedback loop is straightforward:

NGOs generate alarming claims → media amplify those claims → public concern rises → politicians respond with mandates and funding → international treaties lock in the response → bureaucracies expand to implement the treaties → bureaucracies fund the NGOs that generated the original claims.

Each actor behaves rationally within its own incentives. Yet the system produces a ratchet effect: authority moves upward and outward, while failure generates demands for more intervention, rather than reconsideration.

Once a cascade achieves sufficient momentum, it can become embedded in treaties and institutions that resist revision. Agencies accumulate staff, budgets, and stakeholders who depend on their continuation. Narratives of indispensability emerge.

This dynamic helps explain why drug prohibition persists despite decades of failure, why development aid continues despite evidence that it can entrench weak governance, and why tobacco harm reduction faces opposition from institutions that claim to exist to reduce smoking-related harm.

B. Bootleggers, Baptists, and Moralized Policy

Bruce Yandle’s “Bootleggers and Baptists” framework helps explain the durability of these arrangements.^[78]

The analogy comes from communities in the American South where Baptist ministers publicly supported Sunday alcohol bans on moral grounds, while bootleggers quietly supported the same laws because they eliminated legal competition. Yandle used the example to show how regulation often persists because it is supported simultaneously by moral advocates (“Baptists”) and economic beneficiaries (“bootleggers”).

These often-tacit alliances can be remarkably resilient. Moral advocates provide legitimacy and public pressure, while economic beneficiaries supply resources and political support.

The pattern recurs across policy domains. In the drug war, law-and-order moralists provide cover for the DEA, prison-guard unions, and criminal organizations that profit from prohibition. In environmental regulation, green NGOs supply moral justification, while traders and speculators profit from restricted markets. In international tax policy, “tax justice” campaigners provide rhetorical cover for high-tax governments and the bureaucracies whose budgets depend on expanding tax administration.

The most damaging feature of this dynamic is the moralization of policy debate. Once an issue is framed as a moral imperative—saving the planet, protecting children, fighting poverty—it becomes difficult to subject the policy to ordinary cost-benefit analysis. Dissent is treated not as disagreement, but as moral failure.

Evidence that challenges the prevailing narrative may be dismissed or ignored. The epistemic conditions necessary for democratic self-correction gradually erode.

C. The Structural Logic of the DoDO System

Across the policy domains examined in Section III, a common structural pattern emerges.

First, moral urgency overrides evidence. Policy debates become framed as moral imperatives, and dissent is stigmatized, rather than debated.

Second, international treaties constrain democratic revision. Once policy commitments are embedded in treaty law and implemented by bureaucracies whose existence depends on the treaty, reversal becomes difficult through ordinary political processes.

Third, commercial interests align with bureaucratic expansion. Industries arise to profit from implementation—compliance consultants, aid contractors, regulatory specialists, and advocacy organizations. These actors become stakeholders in the continuation of the system.

Fourth, bipartisan consensus can entrench policy. When both major parties support a framework, it becomes institutional infrastructure—rarely questioned and largely insulated from electoral change.

Fifth, failure often justifies expansion. The deeper the crisis appears, the more resources flow toward the institutions managing it. The DEA’s budget grows as prohibition persists. Aid agencies seek more funding when poverty remains. Environmental NGOs mobilize most effectively when new species are declared endangered.

What makes this system parasitic—in the sense suggested by the lancet liver fluke—is not extraction alone. All governance extracts resources. The defining feature is that the system suppresses the host’s defenses. The DoDO complex does not abolish democratic institutions; it alters how they function.

Policies become moralized, rather than debated. Evidence is curated, rather than challenged. Dissent is stigmatized. Failure becomes justification for further expansion.

D. Interrupting the DoDO Life Cycle

The remedy is not to abolish the organizations that comprise the DoDO ecosystem. Many serve legitimate purposes when operating within democratic constraints. The task is to interrupt the self-reinforcing dynamics that allow them to escape those constraints.

Several reforms follow directly from the institutional analysis above.

First, restore subsidiarity by returning authority to the level closest to those affected.

Second, reassert democratic control over international commitments through sunset clauses, periodic reauthorization, and explicit legislative approval.

Third, end the practice of governments funding NGOs that lobby those same governments for expanded programs.

Fourth, increase transparency and contestability in treaty negotiations and intergovernmental decision making.

Finally, re-anchor international cooperation in voluntarism, experimentation, and jurisdictional competition, drawing on the insights of Charles Tiebout, Elinor Ostrom, and the authors of “The Federalist Papers.”^[79] Competition among jurisdictions, local experimentation, and the ability of citizens and firms to exit poorly governed systems all serve as checks on centralized authority.

Biological parasites thrive when host systems lose the ability to defend themselves. Institutional ecosystems behave similarly. The remaining papers in this series will examine the case studies introduced here in greater detail and develop concrete reform proposals tailored to each domain.

The underlying principle is simple: restore democratic accountability over the agents to whom power has been delegated.

The guardians must once again answer to the guarded.

^[1] John Dalberg-Acton, 1st Baron Acton, Letter to Bishop Mandell Creighton (Apr. 5, 1887), reprinted in John Neville Figgis & Reginald Vere Laurence eds., Historical Essays and Studies 504 (Macmillan 1907).

^[2] See Javier Corrales, Autocratic Legalism in Venezuela, 26 J. Democracy 37, 37–51 (2015).

^[3] The term “democratic deficit” was first applied to European governance by David Marquand. David Marquand, Parliament for Europe (Jonathan Cape 1979). For a rigorous elaboration, see Andreas Follesdal & Simon Hix, Why There Is a Democratic Deficit in the EU: A Response to Majone and Moravcsik, 44 J. Common Mkt. Stud. 533 (2006). Robert A. Dahl argues more broadly that international organizations are unlikely to satisfy democratic criteria. See Robert A. Dahl, Can International Organizations Be Democratic? A Skeptic’s View, in Democracy’s Edges 19–36 (Ian Shapiro & Casiano Hacker-Cordón eds., Cambridge Univ. Press 1999).

^[4] P.J. is no longer with us. I hope the analysis that follows honors his spirit.

^[5] The line appears in a longer version of Juvenal’s sixth satire, discovered in an Oxford manuscript by E. O. Winstedt in 1899. See Juvenal, Satires VI, ll. 347–48 (Oxford, Bodleian Library, MS Canonicianus Lat. 41).

^[6] See James M. Buchanan & Gordon Tullock, The Calculus of Consent: Logical Foundations of Constitutional Democracy 3–5 (Univ. of Michigan Press 1962); Friedrich A. Hayek, The Constitution of Liberty 153–63 (Univ. of Chicago Press 1960).

^[7] Montesquieu, The Spirit of the Laws bk. XI, ch. 6 (Thomas Nugent trans., 1750) (1748).

^[8] See M.J.C. Vile, Constitutionalism and the Separation of Powers 83–106 (2d ed., Liberty Fund 1998) (discussing Montesquieu’s idealized account of the English constitution).

^[9] James Madison, The Federalist No. 47, N.Y. Indep. J., Jan. 30, 1788, reprinted in The Federalist Papers 299–306 (Clinton Rossiter ed., Signet Classic 2003) (1788).

^[10] James Madison, The Federalist No. 51, N.Y. Indep. J., Feb. 6, 1788, reprinted in The Federalist Papers 317–22 (Clinton Rossiter ed., Signet Classic 2003) (1788).

^[11] U.S. Const. arts. I–III.

^[12] James Madison, The Federalist No. 39, N.Y. Indep. J., Jan. 16, 1788, reprinted in The Federalist Papers 236–43 (Clinton Rossiter ed., Signet Classic 2003).

^[13] James Madison, The Federalist No. 10, Daily Advertiser (N.Y.), Nov. 22, 1787, reprinted in The Federalist Papers 71–79 (Clinton Rossiter ed., Signet Classic 2003).

^[14] James Madison, The Federalist No. 45, N.Y. Indep. J., Jan. 26, 1788, reprinted in The Federalist Papers 288–93 (Clinton Rossiter ed., Signet Classic 2003).

^[15] U.S. Const. amend. X.

^[16] Charles M. Tiebout, A Pure Theory of Local Expenditures, 64 J. Pol. Econ. 416 (1956).

^[17] Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action 58–69, 149–59 (Cambridge Univ. Press 1990).

^[18] See Friedrich A. Hayek, The Use of Knowledge in Society, 35 Am. Econ. Rev. 519 (1945).

^[19] Pope Pius XI, Quadragesimo Anno: On Reconstruction of the Social Order ¶ 79 (May 15, 1931). The principle builds on Pope Leo XIII, Rerum Novarum: On Capital and Labor (May 15, 1891).

^[20] Treaty on European Union art. 5(3), Feb. 7, 1992, 1992 O.J. (C 191) 1 (entered into force Nov. 1, 1993).

^[21] See Giandomenico Majone, Dilemmas of European Integration: The Ambiguities and Pitfalls of Integration by Stealth 131–38 (Oxford Univ. Press 2005).

^[22] Marquand, supra note 3.

^[23] Aristotle, Politics bk. III, pt. xvi, at 1287a (Benjamin Jowett trans., Oxford Univ. Press 1920).

^[24] Andrew Lintott, The Constitution of the Roman Republic 8–13 (Oxford Univ. Press 1999).

^[25] Alan Watson, The Spirit of Roman Law 44–52 (Univ. of Ga. Press 1995).

^[26] Lintott, supra note 24; Tony Honoré, Emperors and Lawyers 1–15 (Oxford Univ. Press 1994).

^[27] Harold J. Berman, Law and Revolution: The Formation of the Western Legal Tradition 333–35 (Harvard Univ. Press 1983); Bruce L. Benson, The Enterprise of Law: Justice Without the State 30–36 (Pac. Research Inst. 1990); Paul R. Milgrom, Douglass C. North & Barry R. Weingast, The Role of Institutions in the Revival of Trade: The Law Merchant, Private Judges, and the Champagne Fairs, 2 Econ. & Pol. 1, 5–9 (1990).

^[28] Magna Carta cl. 39 (1215) (“No free man shall be seized or imprisoned, or stripped of his rights or possessions … except by the lawful judgment of his equals or by the law of the land.”).

^[29] John Locke, Two Treatises of Government bk. II, ch. XI, § 137 (Peter Laslett ed., Cambridge Univ. Press 1988) (1689).

^[30] Friedrich A. Hayek, The Road to Serfdom ch. 6 (Routledge 1944) (“Planning and the Rule of Law”).

^[31] Hayek, supra note 6, at 205–66.

^[32] The modern formalization of the principal–agent problem begins with Michael C. Jensen & William H. Meckling, Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure, 3 J. Fin. Econ. 305 (1976). For its application to politics, see generally Jean-Jacques Laffont & David Martimort, The Theory of Incentives: The Principal-Agent Model (Princeton Univ. Press 2002).

^[33] See John Ferejohn, Incumbent Performance and Electoral Control, 50 Pub. Choice 5 (1986); see also James D. Fearon, Electoral Accountability and the Control of Politicians: Selecting Good Types Versus Sanctioning Poor Performance, in Democracy, Accountability, and Representation 55–97 (Adam Przeworski et al. eds., Cambridge Univ. Press 1999).

^[34] See Anthony Downs, An Economic Theory of Democracy 207–19 (Harper & Row 1957); Bryan Caplan, The Myth of the Rational Voter: Why Democracies Choose Bad Policies 50–93 (Princeton Univ. Press 2007).

^[35] See generally Adam Przeworski et al., Democracy and Development: Political Institutions and Well-Being in the World, 1950–1990 (Cambridge Univ. Press 2000).

^[36] Alexis de Tocqueville, Democracy in America vol. 1, ch. 11, at 180–92 (Henry Reeve trans., 1835).

^[37] John Stuart Mill, On Liberty ch. 2 (1859), reprinted in On Liberty and Other Essays 20–61 (John Gray ed., Oxford Univ. Press 1991).

^[38] Whitney v. California, 274 U.S. 357, 375–76 (1927) (Brandeis, J., concurring).

^[39] See generally Buchanan & Tullock, supra note 6, at 283–91 (discussing the tendency of bureaucracies to expand beyond their original mandates); William A. Niskanen, Jr., Bureaucracy and Representative Government 36–46 (Aldine Atherton 1971) (presenting the budget-maximizing model of bureaucratic behavior).

^[40] See Kenneth Anderson, After the End of the World: The Decline of International Organizations and the Rise of Global Regulatory Networks, 31 Harv. J.L. & Pub. Pol’y 1, 14–22 (2008).

^[41] Niskanen, supra note 39, at 36–46; see also Mancur Olson, The Logic of Collective Action: Public Goods and the Theory of Groups 5–65 (Harvard Univ. Press 1965).

^[42] See George J. Stigler, The Theory of Economic Regulation, 2 Bell J. Econ. & Mgmt. Sci. 3 (1971) (arguing that regulated industries often capture their regulators); Bruce Yandle, Bootleggers and Baptists: The Education of a Regulatory Economist, 7 Regulation 12 (1983) (describing coalitions between moral advocates and economic beneficiaries of regulation).

^[43] Angus Deaton, The Great Escape: Health, Wealth, and the Origins of Inequality 282–90 (Princeton Univ. Press 2013); Dambisa Moyo, Dead Aid: Why Aid Is Not Working and How There Is a Better Way for Africa 38–52 (Farrar, Straus & Giroux 2009); Andrew Mwenda, Foreign Aid and the Weakening of Democratic Accountability in Uganda, Foreign Pol’y Briefing No. 88, at 5–6 (Cato Inst. 2006); Anne O. Krueger, Aid in the Development Process, 1 World Bank Res. Observer 57, 67–73 (1986).

^[44] Raghuram G. Rajan & Arvind Subramanian, Aid and Growth: What Does the Cross-Country Evidence Really Show?, 90 Rev. Econ. & Stat. 643, 657–60 (2008); Simeon Djankov, José G. Montalvo & Marta Reynal-Querol, The Curse of Aid, 13 J. Econ. Growth 169, 171–74 (2008).

^[45] Craig Burnside & David Dollar, Aid, Policies, and Growth, 90 Am. Econ. Rev. 847 (2000); Rajan & Subramanian, supra note 44; Tseday Jemaneh Mekasha & Finn Tarp, A Meta-Analysis of Aid Effectiveness: Revisiting the Evidence, 7 Pol. & Governance 5 (2019).

^[46] U.S. Agency for Int’l Dev., Localizing Foreign Assistance: Why It Matters and How to Do It 5–7 (2014); Publish What You Fund, U.S. Foreign Aid Transparency and Local Spending Analysis 14–18 (2020).

^[47] Int’l Pol’y Network, Fake Aid: How Foreign Aid Is Being Used to Support the Self-Serving Political Activities of NGOs (2005); see also Christopher J. Coyne, Doing Bad by Doing Good: Why Humanitarian Action Fails (Stanford Univ. Press 2013).

^[48] William Easterly, The White Man’s Burden: Why the West’s Efforts to Aid the Rest Have Done So Much Ill and So Little Good 38–52 (Penguin Press 2006); James Ferguson, The Anti-Politics Machine: “Development,” Depoliticization, and Bureaucratic Power in Lesotho 251–56 (Univ. of Minn. Press 1994).

^[49] Peter J. Haas, Introduction: Epistemic Communities and International Policy Coordination, 46 Int’l Org. 1, 3–5 (1992); Michele M. Betsill & Elisabeth Corell, NGO Diplomacy: The Influence of Nongovernmental Organizations in International Environmental Negotiations, 27 Global Env’t Pol. 65, 67–70 (2001).

^[50] Michael ’t Sas-Rolfes, African Wildlife Conservation and the Evolution of Hunting Institutions, in Wildlife Conservation by Sustainable Use 76, 79–82 (H. H. T. Prins, J. G. Grootenhuis & T. T. Dolan eds., 2000); Brian Child, Sustainable Governance of Wildlife and Community-Based Natural Resource Management: From Economic Principles to Practical Governance 27–34 (Routledge 2019).

^[51] Henry I. Miller & Gregory Conko, The Frankenfood Myth: How Protest and Politics Threaten the Biotech Revolution 119–25 (Praeger 2004); Matin Qaim, Genetically Modified Crops and Agricultural Development 48–52 (Palgrave Macmillan 2016).

^[52] Jonathan H. Adler, The Perverse Effects of Environmental Regulation, 29 Envtl. L. 1, 8–12 (1999); Betsill & Corell, supra note 49, at 67–70; Todd J. Zywicki, Environmental Externalities and Political Externalities: The Political Economy of Environmental Regulation, 73 Tul. L. Rev. 845 (1999); Yandle, supra note 42.

^[53] Jeremy Youde, Global Health Governance 44–56 (Polity Press 2012); Devi Sridhar, Governing Global Health: Who Runs the World and Why? 35–49 (Oxford Univ. Press 2017).

^[54] David T. Levy, Ron Borland, K. Michael Cummings et al., The Global Tobacco Epidemic and the Bloomberg Initiative, 23 Tobacco Control i1 (2014).

^[55] World Health Org., Electronic Nicotine Delivery Systems and Electronic Non-Nicotine Delivery Systems (ENDS/ENNDS): Report by the Secretariat, WHO Doc. FCTC/COP/7/11 (Aug. 9, 2016); World Health Org., WHO Report on the Global Tobacco Epidemic 2021: Addressing New and Emerging Products (2021); David T. Levy et al., The Public Health Impact of E-Cigarettes: A Systematic Review, 25 Tobacco Control i3 (2016).

^[56] The Royal College of Physicians concluded in 2016 that e-cigarettes are unlikely to exceed 5% of the harm associated with smoking. Royal Coll. of Physicians, Nicotine Without Smoke: Tobacco Harm Reduction 87–94 (Apr. 28, 2016). Public Health England reached a similar conclusion, estimating that vaping is at least 95% less harmful and contributes to roughly 20,000 successful quit attempts annually in England alone. See also Pub. Health Eng., E-Cigarettes and Heated Tobacco Products: Evidence Review 12–18 (2018).

^[57] David T. Levy et al., The Potential Impact of E-Cigarettes on Smoking Prevalence and Smoking-Attributable Deaths in the United States, 116 Am. J. Pub. Health 1529 (2017); Jonathan H. Adler & Ilya Shapiro, The FDA’s Misguided War on Vaping, 38 Regulation 14 (2015).

^[58] Devi Sridhar, Who Runs Global Health?, 93 Foreign Affs. 106 (2014); Jeremy Shiffman, Four Challenges That Global Health Networks Face, 390 Lancet 1998 (2017).

^[59] Single Convention on Narcotic Drugs, Mar. 30, 1961, 520 U.N.T.S. 151; David R. Bewley-Taylor, The United States and International Drug Control, 1909–1997 201–04 (Pinter 2001).

^[60] Wendy Sawyer & Peter Wagner, Mass Incarceration: The Whole Pie 2025, Prison Pol’y Initiative (2025); Council on Foreign Relations, Mexico’s Long War: Drugs, Crime, and the Cartels (2026); Inst. for Just., Policing for Profit: The Abuse of Civil Asset Forfeiture (3d ed. 2020); Radley Balko, Rise of the Warrior Cop: The Militarization of America’s Police Forces (2013); Global Comm’n on Drug Pol’y, War on Drugs: Report of the Global Commission 12–18 (2011).

^[61] Peter Alldridge, Money Laundering Law 3–6 (Hart Publ’g 2003).

^[62] See Mark T. Williams, Anti-Money Laundering Effectiveness, Inst. for Int’l Econ. Pol’y Working Paper No. 2019-5 (George Washington Univ. 2019).

^[63] Organisation for Econ. Co-operation & Dev. (OECD), Harmful Tax Competition: An Emerging Global Issue 7–12 (1998).

^[64] Åsa Johansson et al., Taxation and Economic Growth, OECD Econ. Dep’t Working Paper No. 620, at 11–14 (2008); Organisation for Econ. Co-operation & Dev., Tax Policy Reform and Economic Growth 18–19 (2010).

^[65] Organisation for Econ. Co-operation & Dev., Corporate Tax Statistics 20–24 (2023); Organisation for Econ. Co-operation & Dev., Revenue Statistics 2023 15–18 (2023); Michael P. Devereux, Rachel Griffith & Alexander Klemm, Corporate Income Tax Reforms and International Tax Competition, 35 Econ. Pol’y 451 (2002); Ruud A. de Mooij & Sjef Ederveen, Taxation and Foreign Direct Investment: A Synthesis of Empirical Research, 10 Int’l Tax & Pub. Fin. 673 (2003).

^[66] Organisation for Econ. Co-operation & Dev., Standard for Automatic Exchange of Financial Account Information in Tax Matters (2014); on the Common Reporting Standard (CRS) and its effects, see generally Organisation for Econ. Co-operation & Dev., Inclusive Framework on BEPS: Progress Report (2023).

^[67] Fin. Action Task Force, History of the FATF (2019); Alldridge, supra note 61, at 3–6; Jason C. Sharman, The Money Laundry: Regulating Criminal Finance in the Global Economy 71–90 (Cornell Univ. Press 2011); Jason C. Sharman, Power and Discourse in Policy Diffusion: Anti–Money Laundering in Developing States, 52 Int’l Stud. Q. 635, 639–45 (2008).

^[68] Tax Just. Network, The State of Tax Justice 2023 (Tax Just. Network 2023); Oxfam, Tax Battles: The Dangerous Global Race to the Bottom on Corporate Tax (2016); Christian Aid, Death and Taxes: The True Toll of Tax Dodging (2008).

^[69] Tax Just. Network, Annual Report 2022 18–23 (2023); Int’l Trade Union Confederation (ITUC), Corporate Tax Avoidance: A Trade Union Guide (2017); Fin. Transparency Coal., Who We Are (2023); Glob. All. for Tax Just., About the Global Alliance for Tax Justice (2023).

^[70] Paul N. Rosenstein-Rodan, Problems of Industrialisation of Eastern and South-Eastern Europe, 53 Econ. J. 202 (1943); W. W. Rostow, The Stages of Economic Growth: A Non-Communist Manifesto (Cambridge Univ. Press 1960); Hollis B. Chenery & Alan M. Strout, Foreign Assistance and Economic Development, 66 Am. Econ. Rev. 679 (1966).

^[71] Peter T. Bauer, Dissent on Development: Studies and Debates in Development Economics 119 (Harvard Univ. Press 1971); Anne O. Krueger, Aid in the Development Process, 1 World Bank Res. Observer 57 (1986).

^[72] International Convention for the Regulation of Whaling, Dec. 2, 1946, 62 Stat. 1716; Kurkpatrick Dorsey, Whales and Nations: Environmental Diplomacy on the High Seas 132–36 (Univ. of Wash. Press 2013); Tim D. Smith, Scaling Fisheries: The Science of Measuring the Effects of Fishing, 1855–1955 163–70 (Cambridge Univ. Press 1994); D. Graham Burnett, The Sounding of the Whale: Science and Cetaceans in the Twentieth Century 286–91 (Univ. of Chi. Press 2012).

^[73] World Health Org. & U.N. Children’s Fund, Declaration of Alma-Ata: International Conference on Primary Health Care, Alma-Ata, USSR (Sept. 6–12, 1978); Julia Walsh & Kenneth S. Warren, Selective Primary Health Care: An Interim Strategy for Disease Control in Developing Countries, 301 New Eng. J. Med. 967 (1979); Marcos Cueto, The Origins of Primary Health Care and Selective Primary Health Care, 94 Am. J. Pub. Health 1864 (2004); Randall M. Packard, A History of Global Health: Interventions into the Lives of Other Peoples 264–72 (Johns Hopkins Univ. Press 2016).

^[74] Freedom House, Freedom in the World 2025: The Human Cost of Democratic Decline (2025).

^[75] See generally Daniel P. Hughes, Jacques Brodeur & Frédéric Thomas, Host Manipulation by Parasites 7–28 (Oxford Univ. Press 2012).

^[76] Timur Kuran & Cass R. Sunstein, Availability Cascades and Risk Regulation, 51 Stan. L. Rev. 683 (1999).

^[77] Id. at 709–15.

^[78] Yandle, supra note 42, at 12–16.

^[79] See generally Tiebout, supra note 16; Ostrom, supra note 17; Madison, The Federalist No. 10, supra note 13.

Abstract

This essay traces the evolution of online consumer protection from the opening of the Internet to commerce in the mid-1990s through to the challenges posed by artificial intelligence, arguing that trust has always been the central problem and that entrepreneurs, driven by competitive pressure and consumer demand, have repeatedly built the solutions.

I. Trust and Online Commerce

Until the mid-1990s, there was only a modest need for “online consumer protection” because the Internet was essentially free from commerce.[1] This was by design: the main U.S. Internet backbone was operated by the National Science Foundation, which applied an acceptable use policy that restricted commercial activity.[2] In 1992, Congress amended the NSF’s statutory authority to permit the Internet to be used for “additional purposes.”[3] Three years later, the NSF shut down NSFNET, after which the acceptable use policy no longer applied as all Internet traffic ran over commercial backbones.[4]

Once Internet commerce was no longer prohibited, entrepreneurs quickly identified ways to make it possible. For any form of commerce, the key challenge is establishing trust between parties. In-person, trust is built by repetition, reputation, reliable identification, and legal protection. Reputation creates incentives for vendors to maintain quality. Reliable identification is fundamental: without it, repetition is impossible, reputation irrelevant, and legal enforcement futile.[5]

The first attempts to solve the trust problem came from private networks such as CompuServe, Prodigy, and America Online, which starting in the mid-1980s created walled gardens with secure dialup connections, limiting purchases to authorized parties and leveraging their brands and partnerships with brick-and-mortar companies.[6] But this walled garden approach limited users to a single network, constraining the scale of commerce. Outside these gardens, traditional trust mechanisms were largely absent. Building the commercial internet required new tools.

A. Identity and Security

The first challenge was enabling consumers to verify that a website is what it claims to be and that communications are secure. The solution, developed by Taher Elgamal’s team at Netscape, was Secure Sockets Layer (“SSL”), which used public key infrastructure (“PKI”) to authenticate websites through digital certificates and asymmetric cryptography.

When you visit a website over HTTPS (now using transport layer security (“TLS”) rather than SSL but the model remains the same), the server presents a digital certificate that contains the site’s public key, its domain name, and crucially, a digital signature from a Certificate Authority (“CA”).[7] The browser, which maintains a list of trusted CA root certificates, can then verify the certificate by tracing it back to a trusted root.

SSL (and later TLS) also encrypted information between browser and server. The padlock icon in the browser URL bar reassured users that their connection was secure.

B. User Ratings as a Means of Enhancing Trust

User ratings are now so ubiquitous as to seem inseparable from the web. The first significant deployment appears to have been the “feedback forum” that Pierre Omidyar introduced on eBay in February 1996. At launch, he included this note:

“Most people are honest. And they mean well. Some people go out of their way to make things right. I’ve heard great stories about the honesty of people here. But some people are dishonest. Or deceptive. This is true here, in the newsgroups, in the classifieds, and right next door. It’s a fact of life. But here, those people can’t hide. We’ll drive them away. Protect others from them. This grand hope depends on your active participation. Become a registered user. Use our feedback forum. Give praise where it is due; make complaints where appropriate.”[8]

The forum became very popular and Omidyar credits it with much of eBay’s success because it enabled users to generate trust with one another.[9] User ratings also became intrinsic to the success of many other web businesses, including Amazon, Uber, and Airbnb.

Online ratings systems are of course far from perfect and sites must continuously invest in innovation not only to improve their effectiveness but also to avoid contamination from fake reviews (whether generated by humans or bots). Examples of such innovations include: verification of purchase, verification of reviewer, proofs of humanity (e.g. Captcha), and AI-based fraud detection.[10]  But the granularity of user ratings and the degree to which they enable preference matching, combined with the continuous improvements that websites and apps must develop and implement arguably makes them superior to other forms of regulation — especially top-down licenses.[11]

C. Credit Cards, Online Security and Trust

Online commerce requires a means of payment. Early attempts at online money relied on centralized accounts, which floundered for the same reason as the walled gardens: the inability to transact outside a closed network.

Credit cards offered a ready-made solution. Card networks needed only to adapt rules already developed for “card not present” (“CNP”) transactions such as telephone sales.

Merchants trust card issuers because they credibly commit to pay as long as the merchant authenticates properly and consumers find card issuers’ zero-liability fraud guarantees attractive for online transactions.

But Internet transactions posed novel problems: issuers and cardholders feared stolen card details, while merchants feared chargebacks from fraudulent use. Enhanced security on both sides was needed. As noted, SSL/TLS provided part of that solution. Payment networks supplemented these first with their own security standards and then in 2004 with PCI DSS.[12]

Another piece of the puzzle was solved in 1999 with the launch of PayPal, whose most important function has been as a payment gateway — a trusted intermediary processing card payments. Many other online payment gateways, including Worldpay, Stripe, and Square/Block, also now act as a trust layer in the online payments ecosystem.[13]

D. Peer-to-Peer Payments, Stablecoins and Escrow

Over the past 20 years, peer-to-peer systems such as PayPal, Venmo, CashApp, Google Pay, and Apple Cash have become increasingly popular for paying for goods and services, with similar services in India, China, Thailand, Brazil, and elsewhere.[14] Stablecoins are beginning to offer similar functionality, especially in jurisdictions with high inflation.

The near-instant payment finality of such systems is advantageous between trusted counterparties. But for transactions with significant counterparty risk — unknown vendors, undelivered goods, uncertain quality — irreversibility becomes a problem.

In the U.S. and Europe, credit card chargebacks addressed this. In China, where credit card penetration was low, Alibaba developed AliPay in 2003 as a micro-escrow service holding funds until receipt is acknowledged. Similar escrow systems might address counterparty risk for real-time payments and stablecoins.

II. The Implications of AI

The increasingly widespread use of sophisticated AIs presents new issues for online consumer protection. As with the emergent Internet in the mid-1990s, the key issue is trust. In this case, there are three related problems: First, can we trust that the material being presented to us is real (or at least not generated by AI)? Second, can websites trust that the entities seeking to interact with them are real humans (or at least bots authorized to act by real humans)? Third, how can either side trust that the counterparty is who they say they are?

A. Digital Content Credentials and Digital Watermarks

One way to ensure that a digital object was made by a human is for the creator to apply a digital content credential. These use cryptographic signatures to create an auditable history of content (audio, video, text, etc.), documenting who and/or what produced it and altered it.

The most significant such project is the Coalition for Content Provenance and Authenticity (C2PA), with over 5,000 members including Adobe, Google, Meta, Microsoft, OpenAI, and Sony. C2PA uses PKI to establish the origin and history of digital content.[15] C2PA digital credentials are embedded in metadata, so are not part of the content itself, but can be identified by software capable of authenticating the claims.[16] The C2PA standard is open source, so can be implemented by anyone without payment (as long as they abide by the requirements of the Creative Commons license).

OpenAI and Google add C2PA metadata to all visual content produced by their models.[17] As such, an unadulterated image produced by their models should be detectable as such by using a browser extension or C2PA’s tool. However, at least at the time of writing adoption of the browser extension seems muted to say the least (the Chrome version reports 1,000 users).[18]

Surveys suggest about three quarters of consumers in the U.S. are concerned about the authenticity of content.[19] Despite this and despite wide membership on the content side, however, so far there is little evidence of uptake of C2PA or other content authentication systems among content producers, distributors or consumers.[20]

One reason for low consumer-side adoption may be that metadata can be stripped when copying, making C2PA in its original form unreliable. To address this, companies have developed digital watermarks embedded imperceptibly in content itself. Google’s SynthID, for example, attaches unique identifiers to pixels in AI-generated images.[21] Meanwhile, the revised C2PA standard adds a digital watermark or “soft binding,” which acts as a bridge between the asset and its original metadata. If the asset’s metadata is stripped, the watermark can be detected by compatible software, which then queries a “Manifest Repository” to restore the original Content Credentials.

Another explanation is that it is early days. When SSL was first introduced, only ecommerce sites used it. Adoption was gradual until 2014, when Cloudflare’s free Universal SSL reportedly doubled it almost overnight. Similar factors — demonstrable benefits, wider availability, and reduced costs — will likely drive adoption of content authenticity standards.[22]

On the other hand, we might simply learn to live with most forms of AI adulteration, much as CGI is taken for granted in film, and synthesizers and electronic manipulation are widely accepted in music. In that case, the role of identifiers such C2PA and AI detection might be limited to the prevention and prosecution of IP theft and impersonation.

B. The Bot Wars

The problem of bots seeking to fool websites goes back to 1997, when spammers manipulated AltaVista’s search rankings by mass-submitting URLs. AltaVista responded by requiring users to recognize characters displayed so that bots could not easily interpret them.[23]

In the early 2000s, at the request of Yahoo!, Louis van Ahn and colleagues at Carnegie Mellon University developed the Completely Automated Public Turing test to tell Humans and Computers Apart (“CAPTCHA”), which required website users to perform tasks that would be easy for humans but impossible for computers — originally recognition of a word displayed in a manner similar to the AltaVista solution.[24] The first commercial implementation of CAPTCHA was the Gausebeck-Levchin test implemented by PayPal in 2001 to prevent bots from creating fraudulent accounts.[25]

Programmers quickly developed algorithms to defeat CAPTCHAs, sparking an arms race. In 2007, von Ahn developed reCAPTCHA, displaying two words — one known to the computer, one unknown. Aggregating user responses improved machine text recognition. In 2009, von Ahn sold reCAPTCHA to Google, which used it for Google Books and other recognition projects. Von Ahn and Google had strong incentives to ensure reCAPTCHA was effective while minimizing friction, because doing so maximized its use and thereby its value.[26][27][28]

The past 20 years has seen an arms race between CAPTCHA systems and AI. In 2018, Intuition Machines launched a competitor, hCaptcha, which requires users to identify a series of objects and is adaptive to user responses, making it highly effective against AI models. hCaptcha also has an inherently privacy-preserving design; in contrast to Google’s reCAPTCHA, its main source of revenue is the sale of labelled data to train AI models.[29] Also, unlike reCAPTCHA, which is merely free for most customers, hCaptcha pays high-volume websites.

Spam email was another bot-related problem. Matthew Prince and Lee Holloway established Project Honey Pot to track spam sources and enable blocking of malicious IP addresses. The project grew to track threats including DDoS attacks and its creators then developed it into a full-fledged solution called Cloudflare.[30]

Cloudflare’s solutions have become increasingly valuable to website hosts. W3Techs estimates that the company is now responsible for between 21 percent of all internet traff ic, up from about 4 percent in 2016, and represents over 80 percent of all reverse proxy server traffic.[31] In 2020, Cloudf lare switched from reCAPTCHA to hCaptcha, dramatically increasing demand for that service,[32] and in 2023 it switched again, this time to its own alternative called Turnstile.[33] Moreover, all the main bot protection services (reCAPTCHA, hCaptcha, and Turnstile) now offer versions that mainly work in the background, utilizing user analytics that enable them to distinguish humans from bots with considerable accuracy and minimal friction.

These systems protect users whose sensitive data would otherwise be at risk. Innovation driven by entrepreneurial insight and dynamic competition has produced a high level of consumer protection online — and these systems will continue to evolve as AI threats grow more sophisticated.

C.   Online Identity

The third concern raised by AI is its ability to mimic human identity. Criminals use generative AI to create fake identities for various frauds — most being modern versions of old scams such as the sham company: establish a fake company with a name similar to a real company, open a bank account, invoice for fictitious goods or services, transfer funds out immediately they are received, and shut down the company and accounts — leaving little trace of the fraudster’s identity.

Payments with instant finality (e.g. blockchain and real-time payments) and generative AI (which enables quick and easy creation of highly plausible fake IDs) have reduced the cost of implementing sham company frauds.

Fraudsters also use generative AI for impersonation of real humans in real time using deepfake voices and/or video. In one notable case, swindlers held a video call featuring deepfakes of the CFO and various colleagues of engineering firm Arup, thereby convincing an employee of the firm in Hong Kong to send $25 million to a sham company.[34]

Both these frauds are fundamentally examples of impersonation or passing off. While modern technology can make them easier and cheaper (and thereby increase the incentives to undertake them), it also offers solutions to such frauds. Multi-factor authentication (“MFA”) is one relatively simple and inexpensive measure that can be taken: requiring counterparties to prove their identity not only through voice or video but by sharing other trustworthy and ideally immutable evidence.

Similarly, in the case of sham company attacks, company registries that verify beneficial ownership can make such frauds more costly — and in principle make it easier to trace the perpetrators.

Unfortunately, standard MFA documents such a utility bills, drivers’ licenses, and passports can be spoofed by AI. These are widely used for “know your customer” (“KYC”) and beneficial ownership checks and, ironically, can expose personal information to fraudsters when emailed without encryption (as is common).

One form of authentication that is more difficult to spoof and reduces the risk of interception is an immutable digital identity, especially when it is linked to a reliable certificate authority (such as a registry that itself relies on MFA). Jurisdictions such as Estonia and Norway that have implemented such digital IDs have seen rates of fraud decline.[35]

A centralized digital ID registry creates a honeypot risk. Estonia addressed this with the X-Road, a decentralized data exchange architecture with multiple registries each containing limited data. Citizens use a PKI-based eID protected by 2FA: entering a PIN on a physical card unlocks a private key that digitally signs a challenge, confirming identity and authorizing access to specific database(s).[36]

Ideally, one would share a limited credential proving a required fact without revealing underlying personal data. Zero-knowledge proofs (ZKPs) enable this: proving a claim is true without disclosing the sensitive information behind it. Self-sovereign IDs such as Iden3 and PolygonID adopt this model, as does the Yivi app in the Netherlands.

D. AI Agents Creating Fake Human Identities

So far, we have focused mainly on ways in which humans might use AI tools to perpetrate and/or prevent fraud — and how cryptographic solutions can help defeat this threat. More worrying, perhaps, is the possibility of AIs masquerading as humans — a persistent theme of dystopian science fiction.

In principle, digital IDs that require biometric MFA (e.g. facial recognition) might overcome this problem, at least with respect to direct authorization. A further set of problems arise, however, in the context of agentic AIs acting on behalf of humans. First, there is the problem of humans trusting their agents to act as intended. Second, there is the problem of agents being trusted by counterparties.

The recent release of Claude Code, Cowork, OpenClaw, and similar products has created considerable excitement about the prospect that each of us will soon have an army of virtual assistants doing our bidding. At the same time, adoption has likely been slower than it would be if we could trust those agents to do what we want them to do – and if counterparties could reliably trust our agents.

The history of travel agents offers insights. Human agents maintain formal supplier relationships through accreditation  bodies such as ARC or IATA, posting bonds, carrying insurance, and assuming liability. If an agent fails to remit payment, suppliers claim against the bond. Accreditation and financial guarantees make suppliers willing to accept bookings without direct customer relationships.

Online travel agencies like Expedia digitized this model, obtaining the same accreditations and financial guarantees. Their bots were trusted not independently but as authorized arms of accredited, bonded entities.

The innovation was not in the trust model — which remained a known, financially accountable intermediary — but in scale and automation. The supplier trusts the intermediary, which manages all downstream risk.

Autonomous AI agents probably will not be trusted as independent actors anytime soon. Rather, they will operate under accountable intermediaries liable to suppliers. The supplier’s question remains: “do I trust this intermediary?”

This also offers insights into one possible solution to the first problem: trust as an emergent property of agents in general. If third party intermediaries act as a layer between AI agents and suppliers of goods and services, they would effectively signal to consumers which agents are trustworthy.

III. Conclusion

This brief survey of the history and current status of online consumer protection shows how entrepreneurs have addressed trust problems relating to online commerce and how they might now address similar problems in relation to AI.

These developments took place within background conditions favorable to decentralized market solutions: enforceable contracts and the removal of NSF barriers to online commerce. Absent these conditions, it is unclear that online commerce would have emerged as it did.

The specific innovations surveyed illustrate how entrepreneurs build and maintain architectures of trust when background conditions permit. If governments maintain those conditions, entrepreneurs can continue to develop trust architectures that support strong online consumer protection in a world of AI.

[1] See, e.g., Barry M. Leiner et al., Brief History of the Internet, Internet Soc’y (1997), https://www.internetsociety.org/internet/history-internet/brief-history-internet.

[2] National Sci. Found., NSFNET Backbone Servs. Acceptable Use Pol’y (June 1992); see also Nat’l Sci. Found. Off. of Inspector Gen., Review of NSFNET (Mar. 23, 1993).

[3] Scientific and Advanced-Technology Act of 1992, Pub. L. No. 102-476, 106 Stat. 2297 (amending the National Science Foundation Act of 1950 to authorize the National Science Foundation to support access to computer networks that “may be used substantially for additional purposes”).

[4] Susan R. Harris & Elise Gerich, Retiring the NSFNET Backbone Serv.: Chronicling the End of an Era, ConneXions (Apr. 1996), at 2–11. NSFNET was decommissioned Apr. 30, 1995.

[5] This dynamic is illustrated by a simple two-player game with the following payoffs: (cooperate, cooperate = 2,2; cooperate, defect = −1,3; defect, cooperate = 3,−1; defect, defect = 0,0). In a one-shot game without collusion, both players defect, yielding a payoff of 0 for each. If the game repeats indefinitely, however, each player has an incentive to cooperate, producing a payoff of 2 for both. See, e.g., Robert Axelrod & William D. Hamilton, The Evolution of Cooperation, 211 Science 1390 (1981).

[6] See, e.g., Shane Greenstein, How the Internet Became Commercial: Innovation, Privatization, and the Birth of a New Network 109–45 (2015) (describing the “walled-garden” model of early online services).

[7] A certificate authority (CA) acts as a trusted third party. Early providers included VeriSign and Thawte (which VeriSign acquired in 1999). Today, the most widely used CAs include Let’s Encrypt, GlobalSign, Sectigo, and GoDaddy. See, e.g., SSL Certificate Authorities Market Share, W3Techs, https://w3techs.com/technologies/history_overview/ssl_certificate.

[8] Pierre Omidyar, Founder’s Letter, eBay (Feb. 26, 1996), https://pages.ebay.co.uk/services/forum/feedback-foundersnote.html.

[9] Interview with Pierre Omidyar, Founder, eBay Inc., at 4:55, On Innovation (The Henry Ford YouTube channel), https://www.youtube.com/watch?v=RKVmsifohgM&t=295s.

[10] See, e.g., Amazon, How Amazon Is Using AI to Detect Fake Product Reviews and Ensure Authentic Customer Feedback (Aug. 27, 2024), https://www.aboutamazon.com/news/policy-news-views/how-ai-spots-fake-reviews-amazon (describing Amazon’s use of verified-purchase labels, large-language-model and graph-neural-network fraud detection, and expert investigators to block hundreds of millions of suspected fake reviews); Michael Luca & Georgios Zervas, Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud, 62 Mgmt. Sci. 3412 (2016) (finding that Yelp’s automated filter flagged about 16% of restaurant reviews as suspicious).

[11] Julian Morris, Consumer Protection in the 21st Century, ICLE Issue Brief 2023-02-24, at 1 (Feb. 24, 2023), https://laweconcenter.org/wp-content/uploads/2023/02/Morris-consumer-protection-in-the-21st-century-2-24-23.pdf.

[12] PCI Sec. Standards Council, https://www.pcisecuritystandards.org (last visited Feb. 19, 2026).

[13] PayPal Editorial Staff, What Is a Payment Gateway and How Does It Work?, PayPal Bus. Res. Ctr. (June 18, 2024), https://www.paypal.com/us/brc/article/what-is-a-payment-gateway.

[14] Julian Morris, Peer-to-Peer and Real-Time Payments: A Primer, ICLE Issue Brief (Aug. 21, 2023), https://laweconcenter.org/wp-content/uploads/2023/08/RTP-primer.pdf.

[15] Coalition for Content Provenance and Authenticity, https://c2pa.org (last visited Feb. 19, 2026).

[16] The Coalition for Content Provenance and Authenticity (C2PA) uses the X.509 digital-certificate standard, the same standard underlying Secure Sockets Layer (SSL) and Transport Layer Security (TLS).

[17] OpenAI, C2PA in ChatGPT Images, OpenAI Help Ctr., https://help.openai.com/en/articles/8912793-c2pa-in-chatgpt-images (last visited Feb. 19, 2026). C2PA metadata embedded in images generated with ChatGPT or OpenAI’s API allows users to verify provenance information—such as whether an image was created using ChatGPT or the DALL·E-3 model—unless the metadata has been removed.

[18] Digimarc Labs, C2PA Content Credentials, Chrome Web Store, https://chromewebstore.google.com/detail/c2pa-content-credentials/mjkaocdlpjmphfkjndocehcdhbigaafp (last visited Feb. 19, 2026). The extension allows users to verify whether images, audio, or video files contain C2PA manifests and to validate those manifests to reveal information about the content’s origin and history.

[19] See, e.g., Adobe, Content Authenticity and AI Survey (2024) (finding that about 77% of U.S. consumers are concerned about distinguishing real from AI-generated content).

[20] The limited adoption of C2PA browser extensions—see note 17—is one indicator. As of mid-2025, no major social-media platform or news outlet had implemented consumer-facing C2PA verification at scale.

[21] See, e.g., Google DeepMind, SynthID, https://deepmind.google/science/synthid (describing a system that embeds imperceptible digital watermarks in AI-generated images, audio, text, and video so that the content can later be identified as machine-generated); Emma Roth, Google’s SynthID Watermark Can Identify AI-Generated Images, The Verge (Aug. 29, 2023), https://www.theverge.com/2023/8/29/23849107/synthid-google-deepmind-ai-image-detector. SynthID embeds invisible digital watermarks directly into AI-generated media so that they remain imperceptible to users but detectable by specialized tools, helping identify whether content was produced by generative AI.

[22] See, e.g., SSL Certificate Authorities: Historical Yearly Trends in Usage Statistics, W3Techs, https://w3techs.com/technologies/history_overview/ssl_certificate; Cloudflare, Introducing Universal SSL (Sept. 29, 2014), https://blog.cloudflare.com/introducing-universal-ssl (noting that Cloudflare doubled the number of HTTPS sites on the internet when it launched free Universal SSL in 2014).

[23] See, e.g., Andrei Z. Broder et al., Syntactic Clustering of the Web, 29 Computer Networks & ISDN Sys. 1157 (1997); GeeTest, History of CAPTCHA—The Origin Story, https://www.geetest.com/en/article/captcha-origin.

[24] Luis von Ahn, Manuel Blum, Nicholas J. Hopper & John Langford, CAPTCHA: Using Hard AI Problems for Security, in Advances in Cryptology—EUROCRYPT 2003 294 (Eli Biham ed., 2003).

[25] Gausebeck-Levchin Test, Golden, https://golden.com/wiki/Gausebeck-Levchin_test-EKMBKKR (last visited Feb. 19, 2026).

[26] Luis von Ahn et al., reCAPTCHA: Human-Based Character Recognition via Web Security Measures, 321 Science 1465 (2008).

[27] Id.; see also Google, Teaching Computers to Read: Google Acquires reCAPTCHA (Sept. 16, 2009), https://googleblog.blogspot.com/2009/09/teaching-computers-to-read-google.html.

[28] What Is CAPTCHA?, IBM Think, https://www.ibm.com/think/topics/captcha (last visited Feb. 19, 2026); Ciarán Daly, “I’m Not a Robot”: Google’s Anti-Robot reCAPTCHA Trains Their Robots to See, AI Bus. (Oct. 25, 2017), https://aibusiness.com/companies/-i-m-not-a-robot-google-s-anti-robot-recaptcha-trains-their-robots-to-see.

[29] Katharine Schwab, Suspicious of Google’s reCAPTCHA? Here’s a Popular Alternative, Fast Company (July 17, 2019), https://www.fastcompany.com/90377406/suspicious-of-googles-recaptcha-heres-a-popular-alternative; Intuition Machines, About hCaptcha, https://www.hcaptcha.com/about.

[30] Cloudflare, Our Story, https://www.cloudflare.com/our-story (describing how Matthew Prince and Lee Holloway founded Project Honey Pot in 2004 to track how spammers harvested email addresses); see also Project Honey Pot, https://www.projecthoneypot.org (last visited Feb. 19, 2026); Matthew Prince, Cloudflare’s First Birthday, Cloudflare Blog (Sept. 27, 2011), https://blog.cloudflare.com/cloudflares-first-birthday.

[31] Historical Yearly Trends in the Usage Statistics of Reverse Proxy Services for Websites, W3Techs, https://w3techs.com/technologies/history_overview/proxy/all/y.

[32] Thomas Claburn, Cloudflare Dumps Google’s reCAPTCHA, Moves to hCaptcha as Free Ride Ends (and Something About Privacy), The Register (Apr. 9, 2020), https://www.theregister.com/2020/04/09/cloudflare_dumps_recaptcha.

[33] Reid Tatoris, Adam Martinetti & Benedikt Wolters, Cloudflare Is Free of CAPTCHAs; Turnstile Is Free for Everyone, Cloudflare Blog (Sept. 28, 2023), https://blog.cloudflare.com/turnstile-ga; Migrate from hCaptcha, Cloudflare Turnstile Docs, https://developers.cloudflare.com/turnstile/migration/hcaptcha (last visited Feb. 19, 2026).

[34] Heather Chen & Kathleen Magramo, Finance Worker Pays Out $25 Million After Video Call with Deepfake ‘Chief Financial Officer’, CNN (Feb. 4, 2024), https://www.cnn.com/2024/02/04/asia/deepfake-cfo-scam-hong-kong-intl-hnk.

[35] See, e.g., Estonian Digital Identity, e-Estonia, https://e-estonia.com/solutions/e-identity/id-card; see also BankID Norge, Norwegian Digitisation Agency, https://www.bankid.no.

[36] X-Road Architecture, X-Road, https://x-road.global/architecture; Republic of Est. Info. Sys. Auth., X-Road Factsheet, https://www.ria.ee/en/state-information-system/x-tee; see also Privacy Int’l, ID Systems Analysed: e-Estonia (2021), https://privacyinternational.org/case-study/4737/id-systems-analysed-e-estonia; Nordic Inst. for Interoperability Sols. (NIIS), X-Road World Map, https://x-road.global/xroad-world-map (showing implementations in Finland, Iceland, Japan, Ukraine, and other jurisdictions); see also Nordic Inst. for Interoperability Sols. (NIIS), About NIIS, https://niis.org.

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).

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.

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.