ICLE Comments to the European Commission on Copyright in the DSM Directive
I. Introduction
The International Center for Law & Economics (ICLE) respectfully submits these comments in response to the European Commission’s call for evidence on the review of the Copyright in the Digital Single Market (CDSM) Directive and the contemplated targeted initiative. ICLE is a nonprofit, nonpartisan research centre dedicated to building the intellectual foundations for sensible, economically grounded policy. ICLE has written extensively on copyright and artificial intelligence (AI), the design of intermediary-liability rules, and copyright enforcement against online piracy. We offer these comments in that spirit.
Our comments focus on two of the four problem areas the Commission has identified: copyright’s application to generative AI training, including the text-and-data-mining (TDM) exceptions and the ‘reservation of rights’ mechanism in Article 4(3) of the CDSM Directive; and copyright enforcement against online piracy, particularly piracy of live and other time-sensitive content.
These are the areas where the underlying economics point most clearly to workable answers—and where counterproductive regulation poses the greatest risk.
A single theme connects our analysis of both issues: copyright is an instrument for aligning incentives, not an end in itself. Rules work best when they place obligations on the party best positioned to act at the lowest cost, and when they channel disputes toward markets and bargaining, rather than rigid prohibitions that suppress beneficial activity.
For AI, that counsels a measured approach that preserves room for voluntary transactions to emerge, rather than freezing today’s standoff between rightsholders and developers into law. For piracy, it counsels targeted, incentive-compatible enforcement—above all, no-fault, dynamic, and ‘live’ blocking injunctions—rather than broad new liability mandates that would push intermediaries toward over-removal of lawful expression.
We take each in turn.
II. Generative AI, Copyright, and the Right Regulatory Margin
The Commission’s review should start from a simple premise: the European Union already has a framework for AI training. The Copyright in the Digital Single Market (CDSM) Directive supplies text-and-data-mining (TDM) exceptions and a mechanism for rightsholders to reserve rights. The Artificial Intelligence Act adds obligations for general-purpose AI (GPAI) model providers to respect those reservations and publish summaries of training content.
The question, then, is not whether to regulate from a blank slate. It is whether to preserve a flexible, opt-out-based framework that can support voluntary licensing and targeted enforcement, or to move toward a rigid input-licensing regime that would raise transaction costs, entrench incumbents, and do little to compensate creators for actual market harm. The better course is to focus policy where value is realised and injury is most cognisable: AI outputs, not inputs.
A. The EU Already Regulates AI Training
The CDSM Directive establishes two TDM exceptions: a mandatory exception for scientific research by research and cultural-heritage institutions under Article 3, and a broader exception for TDM by any party under Article 4. The Article 4 exception applies unless the rightsholder has expressly reserved its rights, including through machine-readable means.[1]
The Commission reads Article 4 as capable of covering commercial AI training on publicly available online content, provided the user has lawful access and the rightsholder has not validly reserved its rights.[2] Within this framework, TDM opt-outs are an essential mechanism through which rightsholders may reserve rights, particularly for AI model training.[3]
The AI Act builds on this architecture. It requires providers of GPAI models to adopt a policy to identify and respect rights reserved under Article 4(3), and to publish a sufficiently detailed summary of the content used for training.[4] The practical effect is that AI training in the EU already operates within a defined opt-out regime, paired with transparency obligations and a developing market for the technical tools needed to express and honour reservations of rights.[5]
That matters for this review. The Commission need not invent a framework from scratch. It should instead decide how best to refine an architecture already in place. Our central submission is that the Commission should resist proposals that would convert the current opt-out, transparency-based regime into something closer to a mandatory opt-in or input-licensing requirement. It should focus instead on the stage where value is realised and harm to creators is most cognisable: AI outputs.
B. Input Licensing Is the Wrong Place to Focus
There is a strong intuition that, because models are trained on creative works, creators should be paid for each work used as an input. That intuition is understandable, but the economics do not support building policy around it. Three problems recur.
First, input-stage valuation is, at least for now, intractable. A modern model is trained on billions of text fragments and images. The marginal contribution of any single work—even a celebrated one—to the model’s behaviour is vanishingly small and effectively impossible to isolate. As we have previously explained, even the collected works of a major author constitute only a minute fraction of a large training corpus. Any per-work fee applied across millions of works will therefore be detached from each work’s actual contribution.[6] Value in generative AI is more likely to be realised at the output stage, when a model produces something a user will pay for—not at ingestion.
Second, an input-licensing or robust reservation-of-rights regime imposes transaction costs that frustrate the bargaining it is supposed to encourage.[7] Much of the content used in training is not administered by any collective body. Developers cannot realistically identify and negotiate with every rightsholder. That friction tends to produce either degraded models, trained on narrower or synthetic data, or entrenchment of the few incumbents able to afford comprehensive licensing. Smaller and independent creators—precisely those least likely to be inside a collective—gain little, while coverage gaps grow.[8]
Third, an input-side veto risks granting rightsholders a power that exceeds copyright’s proper scope. Copyright protects expression, not ideas or statistical patterns. A model that learns the distribution of language from a work, without reproducing its protected expression in outputs, is doing something different from copying.[9] Treating analysis itself as the locus of liability would expand the copyright entitlement well beyond its traditional bounds, with little corresponding benefit to creators.
The Commission should also be wary of mandated collective licensing. Voluntary collective arrangements can reduce transaction costs, but compulsory ones can create monopolistic structures that warrant competition scrutiny.[10] Where licensing markets can develop voluntarily—and several major content owners and AI developers have already struck direct deals—policymakers should let them, intervening only on evidence of genuine market failure.[11]
C. Focus on Outputs, Transparency, and New Revenue Streams
The more promising path is to focus protection and remuneration on outputs. Creators should have a clear, low-friction mechanism to challenge AI outputs that are substantially similar to their protected works. That approach maps onto U.S. copyright law’s existing substantial-similarity analysis, rather than expanding the entitlement at the input stage.[12] Output-focused enforcement targets the harm creators actually experience—market-substituting reproductions—while preserving room for genuinely transformative uses that serve different markets than the originals.[13]
An output focus also opens new revenue channels that input licensing would foreclose. A workable model could resemble a content-identification system that pays a creator when a user generates material in that creator’s identifiable style. It could also let performers and artists bargain over the use of their name, image, and likeness at the generation stage.[14] These mechanisms tie compensation to actual, commercially valuable use, rather than speculative input value. They are more likely to produce durable bargains between creators and developers.
The Commission’s interest in ‘new revenue streams’ and in strengthening protections for performers against AI-generated impersonations points in exactly this direction. Any new remuneration mechanisms should therefore focus on outputs and identity, not ingestion.
Two implementation cautions follow. First, the Article 53 ‘sufficiently detailed summary’ can support a healthy licensing ecosystem, but only if the disclosure obligation remains proportionate. Overly granular disclosure requirements can be weaponised in litigation and can disadvantage the more transparent developers, without commensurate benefit to creators. The template should enable rights reservation and good-faith licensing, not manufacture infringement exposure.[15]
Second, the CDSM review will interact with the AI Act’s ‘systemic risk’ framework and GPAI obligations. We have cautioned that vague, open-ended designations invite expansive regulatory intervention that can deter AI development in the EU without evidence of tangible harm.[16] A copyright framework that is rigid on inputs, layered atop an expansive AI Act risk regime, would compound the competitiveness concerns already voiced across Europe.
The objective should be a framework flexible enough to let mutually beneficial markets form. That outcome ultimately serves creators better than a brittle entitlement that entrenches the present standoff.[17]
III. Targeted Enforcement Against Online and Live-Event Piracy
The Commission rightly identifies online piracy—especially piracy of live sport and other time-sensitive content—as a persistent threat for which existing remedies have proved insufficient. The core problem is structural. Notice-and-takedown regimes were designed for a slower internet. They operate reactively, creating an endless game of ‘whack-a-mole’ in which infringing material reappears as quickly as it is removed.[18]
For live content, the mismatch is acute. A pirated broadcast’s value is concentrated in the minutes it is live, so a remedy that takes hours or days to act protects little. Illicit IPTV services and sophisticated streaming apps have professionalised this activity, often operating from outside the jurisdiction and beyond the practical reach of ordinary litigation.
A. No-Fault Blocking Is the Best-Fit Remedy
One of the most effective tools to emerge internationally is the no-fault injunction. Under this model, a court or comparable supervised body directs an intermediary, such as an internet service provider (ISP), to disable access to a site or stream determined to be infringing, without any finding that the intermediary itself is at fault.[19]
This is incentive-compatible regulation. It places the obligation on the party best positioned to act at the lowest cost, leverages capabilities intermediaries already have, and—critically—does not impose new primary or secondary liability or any general duty to monitor. Intermediaries gain legal certainty and immunity for good-faith compliance. Rightsholders gain an enforcement mechanism that can keep pace with infringement.
The EU already supplies the legal foundation. Article 8(3) of the InfoSoc Directive requires member states to allow injunctions against intermediaries whose services are used to infringe.[20] Several member states have built effective, well-supervised systems on that authority. U.K. courts moved from static blocking to dynamic and then to ‘live’ blocking orders that operate in real time during sports broadcasts.[21] Germany’s industry-run Clearing Body for Copyright on the Internet (CUII), operating under the oversight of the Federal Network Agency, pairs speed with procedural safeguards and unanimous review before any block.[22]
These models, together with comparable frameworks in Italy, Denmark, and elsewhere, show that fragmentation across member states is itself a weakness the review can address.
The empirical record is encouraging. Studies of website blocking find meaningful reductions in piracy and diversion of users toward legal sources, with limited circumvention in practice. Blocking of a single major site has been associated with double-digit reductions in infringing activity, even among users who used other sites.[23] Blocking also lowers enforcement costs relative to suing individual infringers, making enforcement accessible to smaller rightsholders priced out of conventional litigation.[24] The European Union Intellectual Property Office’s (EUIPO) own study of dynamic blocking injunctions documents their growing and effective use across the Union.[25]
B. Safeguards Should Keep Blocking Narrow and Accountable
Blocking is a powerful tool, and its legitimacy depends on careful design. Poorly calibrated regimes can overblock lawful content, raising free-expression concerns and chilling the creativity copyright is meant to foster. Liability-based mandates that push intermediaries toward invasive filtering carry high private and social costs.[26]
The answer is not to forgo blocking, but to structure it well. Drawing on international experience, an effective EU framework should:
- Define clear, evidence-based criteria for identifying genuinely infringing ‘rogue’ sites and streams.
- Provide streamlined, transparent procedures with notice and a meaningful opportunity for affected sites, intermediaries, and interested third parties to be heard and to seek review.
- Authorise dynamic and live orders, so rightsholders need not return to court for every new domain, IP address, or stream used to evade a block.
- Confine obligations to narrowly tailored, technically feasible measures.
- Shield compliant intermediaries from liability and provide reasonable cost recovery.[27]
Two further points follow.
First, ISP-level blocking is a sound starting point, but the relevant infrastructure also includes content-delivery networks, hosting providers, and domain registrars. The framework should be able to reach other well-placed intermediaries where appropriate, subject to the same safeguards.
Second, blocking complements rather than replaces the broader bargain underlying intermediary-liability law. Safe harbours should protect intermediaries that act responsibly, while ensuring that known, repeated, large-scale infringement does not receive practical tolerance by default.[28]
Properly designed, no-fault blocking aligns the interests of courts, rightsholders, and intermediaries without sacrificing the open internet.
IV. Conclusion
Across both issues, the economically sound path is the same: regulate at the margin where behaviour can actually be influenced at the lowest cost, and prefer markets, bargaining, and targeted remedies to broad prohibitions. Copyright should align incentives, not harden today’s disputes into rigid entitlements that deter innovation or chill lawful expression.
For AI, that means preserving the EU’s existing opt-out and transparency architecture, rather than converting it into an input-licensing mandate. The Commission should allow voluntary licensing markets to develop where they can, keep disclosure obligations proportionate, and channel remuneration toward outputs and identity-based rights, where value is realised and new markets can genuinely form.
For piracy, that means building on Article 8(3) to establish well-safeguarded no-fault, dynamic, and live blocking injunctions. Such tools can target bad actors at speed, reduce enforcement costs, and preserve lawful expression, while avoiding broad liability mandates that would push intermediaries toward over-removal.
The Commission should refine the framework it already has, not replace it with brittle rules that entrench incumbents, raise transaction costs, or weaken Europe’s competitiveness. ICLE thanks the Commission for its consideration and welcomes the opportunity to provide further detail.
[1] Directive (EU) 2019/790 of the Eur. Parliament and of the Council of 17 Apr. 2019 on Copyright and Related Rights in the Digital Single Market and Amending Directives 96/9/EC and 2001/29/EC arts. 3–4, 2019 O.J. (L 130) 92 (text-and-data-mining exceptions).
[2] Eur. Comm’n, Call for Tenders for a Feasibility Study on a Central Registry of Opt-Outs Under the Text and Data Mining (TDM) Exception (2025), https://digital-strategy.ec.europa.eu/en/funding/call-tenders-feasibility-study-central-registry-opt-outs-under-text-and-data-mining-tdm-exception.
[3] Id.
[4] Kristian Stout, Systemic Risk and Copyright in the EU AI Act, Truth on the Mkt. (19 Mar. 2024), https://truthonthemarket.com/2024/03/19/systemic-risk-and-copyright-in-the-eu-ai-act.
[5] Regulation (EU) 2024/1689 of the Eur. Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) art. 53(1)(c)–(d), 2024 O.J. (L) 1 (requiring providers of general-purpose AI models to adopt a policy respecting rights reserved under art. 4(3) of Directive (EU) 2019/790 and to publish a sufficiently detailed summary of training content).
[6] Kristian Stout, ICLE Comments to OSTP on Development of an AI Action Plan (Int’l Ctr. for L. & Econ., 14 Mar. 2025), https://laweconcenter.org/resources/icle-comments-to-ostp-on-development-of-an-ai-action-plan.
[7] See Richard A. Posner, Economic Analysis of Law 42 (7th ed. 2007) (explaining that copyright transaction costs include identifying rightsholders and negotiating licences).
[8] Ben Sperry & Kristian Stout, ICLE Comments to the UK Intellectual Property Office Copyright and AI Consultation (Int’l Ctr. for L. & Econ., 25 Feb. 2025), https://laweconcenter.org/resources/icle-comments-to-the-uk-intellectual-property-office-copyright-and-ai-consultation.
[9] Id.
[10] See, e.g., Broad. Music, Inc. v. CBS, Inc., 441 U.S. 1 (1979) (holding that even voluntary collective-licensing arrangements warrant antitrust scrutiny).
[11] Kristian Stout, Ben Sperry & Subiksha Ramakrishnan, Generative AI: When Fair Use Becomes Unfair Competition, 33 Geo. Mason L. Rev. (forthcoming 2026), https://laweconcenter.org/resources/generative-ai-when-fair-use-becomes-unfair-competition.
[12] Sperry & Stout, supra note 8. See also Giuseppe Colangelo, A Competition Policy Analysis of Copyright Protection in Generative AI, Sing. J. Legal Stud. 271 (2025).
[13] Stout, Sperry & Ramakrishnan, supra note 11.
[14] Kristian Stout, Geoffrey A. Manne & Emily Corbeille, ICLE Comments on Artificial Intelligence and Copyright (Int’l Ctr. for L. & Econ., 30 Oct. 2023), https://laweconcenter.org/resources/icle-comments-on-artificial-intelligence-and-copyright.
[15] Stout, supra note 4.
[16] Id.
[17] Stout, Sperry & Ramakrishnan, supra note 11.
[18] Kristian Stout & Geoffrey A. Manne, A Roadmap to Reform Section 512 of the Copyright Act (Int’l Ctr. for L. & Econ., 13 Oct. 2022), https://laweconcenter.org/resources/a-roadmap-to-reform-section-512-of-the-copyright-act.
[19] Kristian Stout, Julian Morris & Subiksha Ramakrishnan, Site Blocking and Incentive-Compatible Solutions to Illicit Online Activity (Int’l Ctr. for L. & Econ., 15 Aug. 2025), https://laweconcenter.org/resources/site-blocking-and-incentive-compatible-solutions-to-illicit-online-activity.
[20] Directive 2001/29/EC of the Eur. Parliament and of the Council of 22 May 2001 on the Harmonisation of Certain Aspects of Copyright and Related Rights in the Information Society art. 8(3), 2001 O.J. (L 167) 10 (authorising rightsholders to seek injunctions against intermediaries whose services are used to infringe copyright).
[21] Football Ass’n Premier League Ltd. v. British Sky Broad. Ltd., [2013] EWHC 2058 (Ch); see also subsequent Football Ass’n Premier League live-blocking orders directed at U.K. internet service providers during match broadcasts.
[22] See Stout, Morris & Ramakrishnan, supra note 19 (discussing Germany’s Clearing Body for Copyright on the Internet (CUII) and the role of the Federal Network Agency (BNetzA)).
[23] Brett Danaher, Michael D. Smith, Jonathan Samuel Hersh & Rahul Telang, The Effect of Piracy Website Blocking on Consumer Behavior, 44 MIS Q. 631 (2020).
[24] Kristian Stout, Canadian Site Blocking Proposal Is a Good Experiment in Controlling Piracy, Truth on the Mkt. (11 June 2018), https://truthonthemarket.com/2018/06/11/canadian-site-blocking-proposal-is-a-good-experiment-in-controlling-piracy.
[25] Eur. Union Intell. Prop. Off. (EUIPO), Study on Dynamic Blocking Injunctions in the European Union (2021).
[26] Stout, Morris & Ramakrishnan, supra note 19.
[27] Id.
[28] Stout & Manne, supra note 18.