Contents

Introduction. 1

Background: The Regulatory Landscape 2

The TDM Exception: Preserving the Training-Data Framework 3

The Training/Output Distinction Is Critical and Must Not Be Collapsed. 5

Training Data Transparency: A Proportionate Middle Ground. 6

Performer Impersonation: A Targeted Rights Problem Requiring A Targeted Solution. 7

Competitiveness Implications for the Transatlantic AI Ecosystem. 8

The Research Exception and Out-Of-Commerce Works: Pro-Innovation Opportunities 8

Online Piracy and Equitable Remuneration. 9

Recommendations 10

Conclusion. 11

Endnotes 11

Introduction

The Information Technology and Innovation Foundation (ITIF; Transparency Register #: 923915716105-08) is pleased to submit these comments in response to the European Commission’s open consultation on the Targeted Initiative for a Better Copyright Environment for European Creativity and Innovation (Initiative No. 18173).[1] ITIF is a U.S.-based nonprofit, nonpartisan public policy think tank committed to articulating and advancing pro-productivity, pro-innovation, and pro-technology policy agendas around the world to spur growth, prosperity, and progress.

ITIF welcomes the Commission’s effort to review the copyright environment in light of rapid advances in generative AI. This initiative arrives at a critical moment: The EU’s regulatory choices on AI training and copyright will have significant and lasting consequences not only for European creativity and innovation, but for the global AI ecosystem. Because the United States remains the leading developer of frontier AI systems and the primary architect of the global AI technology stack, the Commission’s decisions will directly affect U.S. companies’ ability to develop and deploy AI in and for European markets.

ITIF’s overarching position is that the Commission should preserve and strengthen the text and data mining (TDM) exception established in the Digital Single Market (DSM) Directive as the foundational framework for AI training, while ensuring that meaningful and enforceable protections exist against output-side harms, i.e., cases where AI systems reproduce or substitute for copyrighted content rather than learn from it. These are legally and functionally distinct problems requiring distinct solutions, and conflating them risks both over-regulating AI development and under-protecting creators where harms are genuine. The Commission should resist broad mandatory licensing for training data while supporting targeted and proportionate responses to demonstrated output-side harms and transparency deficits. The European Parliament’s own resolution on copyright and generative AI, adopted on 10 March 2026 by 460 votes to 71, similarly distinguished training-side and output-side questions and called for clarification of licensing rules rather than abandonment of the TDM exception.[2] This is a framing that ITIF endorses.[3]

Background: The Regulatory Landscape

The DSM Directive of 2019 introduced a general text and data mining exception in Article 4, which permits TDM of lawfully accessible content for any purpose, subject to a rightsholder opt-out expressed through machine-readable means. The EU AI Act, adopted in 2024, operationalized this framework for general-purpose AI (GPAI) model providers, requiring them to implement policies to comply with EU copyright law and to publish summaries of training datasets. The General-Purpose AI Code of Practice (GPAI CoP) details how providers can demonstrate compliance, and the Commission’s December 2025 consultation on TDM opt-out protocols advances that implementation.[4]

This Targeted Initiative appears to contemplate further legislative intervention, potentially including mandatory remuneration mechanisms, expanded opt-out rights, or new rules on AI-generated works. ITIF has closely observed analogous deliberations in the United Kingdom, where it has argued for a broad, permissive TDM exception and cautioned against measures that would stifle AI development without commensurate benefits to creators.[5] The same analysis applies to the EU, with the added consideration that EU regulatory choices carry greater global weight given the scale and influence of the European single market.

At the same time, ITIF recognizes that rightsholders have raised legitimate concerns that deserve serious engagement: the use of AI systems to reproduce or closely mimic copyrighted content in outputs, the practical barriers small creators face in exercising opt-out rights, the difficulty in independently determining when the creation or use of an AI system violates copyright of specific works, and the unsettled state of fair use and TDM doctrine in multiple jurisdictions. These concerns are not resolved simply by defending the existing TDM framework, and the Commission’s initiative is an appropriate venue in which to address them.

The Tdm Exception: Preserving the Training-Data Framework

The ability to train AI models on large datasets of text, images, audio, and other content is a technical prerequisite for AI development. Restricting access to training data, whether through mandatory compensation, expanded opt-out rights, or onerous compliance requirements, can have the effect of restricting AI development itself.

Within the above context, the EU’s existing TDM exception under Article 4 of the DSM Directive is a sound foundation that reflects a reasonable balance: developers may train on lawfully accessible content; rightsholders who object may opt out through machine-readable means; and the exception does not authorize reproduction or redistribution of copyrighted works in a manner that substitutes for the original.

A. Mandatory Training-Data Licensing Would Harm Innovation Without Helping Creators

Calls for mandatory licensing or remuneration schemes, under which AI developers would be required to pay collective compensation to rightsholders whose works appear in training datasets, are appealing but would be counterproductive in practice. ITIF offers several reasons for this conclusion:

1. Unworkable at scale. Training datasets for frontier AI models contain hundreds of billions of words and millions of images sourced from the open web, digitized archives, and licensed corpora. No technically feasible mechanism for identifying, attributing, and compensating individual rightsholders proportionally to each work’s contribution to a trained model currently exists, and is unlikely to exist for retroactive application.

2. Chilling effect on European AI development. Broad and sweeping mandatory licensing would substantially increase compliance costs and legal uncertainty for AI development in the EU. Given that AI developers have access to global compute infrastructure and can train models outside the EU, the most likely practical response would be to avoid EU-sourced data, which would then likely have the impact of reducing the representation of European languages, cultures, and perspectives in frontier AI systems.

3. Benefits would accrue to aggregators, not individual creators. In practice, licensing revenues would flow primarily to large collective management organizations and major media conglomerates. The long tail of individual creators—the very constituency the Commission seeks to protect—would see little direct benefit while facing increased barriers to participating in AI-driven creative tools.

4. The voluntary licensing market is beginning to emerge. AI developers are actively negotiating bilateral data licensing deals with news publishers, book publishers, and music rights holders. Imposing a broad mandatory collective scheme would crowd out this nascent market, potentially remove the incentive for bilateral negotiation, and substitute a bureaucratic allocation mechanism for market-determined terms. The existence of these deals also suggests that the voluntary market can function without mandates.

5. The copyright debate is bigger than training data. The narrow focus on training datasets ignores the reality of how AI systems may use copyrighted works. For example, an AI system may use retrieval-augmented generation (RAG) to incorporate copyright materials not in the training phase but during inference to acquire authoritative data. Moreover, AI developers are interested in licensing copyrighted works not only for training, but for other purposes, such as to legally provide direct access to this material to their users or allow users to legally remix copyrighted content. Mandatory training-data licensing only addresses a narrow issue, leaving many bigger ones untouched.

ITIF acknowledges, however, that the voluntary market argument cuts in more than one direction: the fact that AI developers are willing to pay for licensed data suggests that high-quality, curated content carries genuine value that the TDM exception allows developers to capture without compensation. This is a reasonable critique, and the Commission may appropriately ask whether the current framework adequately reflects the contribution of European creative content to the value of AI systems trained on it. ITIF’s view is that the answer lies in robust voluntary markets and targeted transparency requirements—not broad mandatory licensing—but recommends that the Commission should evaluate this tension directly in its impact assessment.

B. The Opt-Out Mechanism Must Remain Technically Workable

While ITIF supports traditional copyright protections, it opposes the EU’s approach of legally restricting AI training inputs under the DSM Directive. Instead, ITIF argues that policymakers should incentivize AI developers to voluntarily respect machine-readable opt-outs by offering them a safe harbour from copyright liability if they do so. It also encourages policymakers to consider enforcement problems faced by smaller creators. Large publishers and professional rights management organizations can engage legal counsel, monitor compliance, and seek relief in court. Individual photographers, independent authors, musicians, and other small creators often do not have the resources to do the same. The Commission should take this asymmetry into account.

ITIF’s recommendation is not to abandon the opt-out framework, which remains the right structural approach, but to ensure that the opt-out infrastructure is genuinely accessible and that compliance obligations are meaningful. Concretely, ITIF recommends that the Commission should:

1. Base required protocols on established, low-cost mechanisms. Robots.txt and HTTP headers represent the current state of the art. These are widely deployed and technically sound. The Commission should adopt them as the primary standard and avoid mandating supplementary mechanisms that create overlapping or conflicting signals.

2. Support voluntary, consensus-driven industry standards for opt-out metadata and tools. The Commission could appropriately support the development of shared technical standards and infrastructure (operated by collecting societies, creator associations, or a public body) that allows individual creators to express opt-out preferences for individual works.

3. Apply opt-out obligations prospectively only. Requiring AI developers to retrain or reconstruct models based on post-hoc assertions of opt-out rights would be technically impossible and would create indefinite legal liability. The Commission should clarify that opt-out obligations are prospective.

4. Keep the scope of opt-out bounded. The opt-out mechanism applies to TDM for AI training. It should not extend to AI inference, to AI-generated outputs, or to the incidental retention of learned statistical patterns in model weights, as those do not constitute “reproductions” of copyrighted works in a legally meaningful sense.

The Training/Output Distinction Is Critical and Must Not Be Collapsed

The single most important analytical distinction the Commission should draw in this initiative is between AI training and AI output. These are different legal and technical questions, and conflating them leads to regulatory responses that are simultaneously over-inclusive (restricting beneficial training uses) and under-inclusive (failing to address genuine output-side harms). ITIF recommends that the Commission structure its initiative around this distinction explicitly.

A. Training-Data Use and Output Reproduction Are Legally Distinct

AI training involves the processing of large corpora to develop statistical representations of language, image, or audio patterns. The trained model does not store or retrieve the training data; it encodes learned relationships across billions of parameters. This is a transformative, computational process that is functionally analogous to human reading and learning, and it is the activity that the DSM Directive’s TDM exception was designed to permit.

Output reproduction is a different matter entirely. When an AI system is used in ways that reproduce, closely paraphrase, or substitute for copyrighted content in its outputs, particularly in real-time retrieval contexts such as AI-powered search or retrieval-augmented generation (RAG), the legal analysis changes. A system that pulls live content from specific websites or databases and delivers it to users in substantially the same form, without directing users to the original source, may be directly displacing the market for that content. This is not a training-data question; it is an output question, and existing copyright law already provides tools to address it.[6]

The Commission should ensure that any new legislative measures are clearly scoped to one or the other context, and that measures designed to address output-side harms are not drafted so broadly as to capture permissible training-data uses as well.

B. Output-Side Harms Are Real and Deserve Targeted and Proportionate Responses

ITIF recognizes that output-side harm to rightsholders is a genuine and growing concern, particularly for news publishers, professional journalists, and other producers of time-sensitive content that AI systems may reproduce or summarize in ways that reduce direct traffic to the original source.

The appropriate response, however, should be targeted and proportionate rather than sweeping. ITIF recommends:

1. Enforce existing copyright against verbatim or near-verbatim output reproduction. Cases in which AI systems reproduce copyrighted content in substantially identical form (whether through RAG, cached content, or direct reproduction) are addressable under the existing copyright framework without new legislation. The Commission should support the enforcement of existing rules rather than layering on new rights that could have unintended effects on training.

2. Distinguish market substitution from transformation. The relevant legal question for output-side reproduction is whether the AI system’s output displaces the market for the original work. Where a summary, paraphrase, or synthesis adds informational value and directs users to the original source, this is less likely to constitute infringing substitution. Where an AI product delivers substantially the same content as the original, without attribution or traffic referral, the harm is more direct. Regulatory or legislative guidance clarifying this distinction would be valuable.

3. Do not extend output-side remedies to training. The Commission should explicitly state that any new measures addressing AI output reproduction do not affect or narrow the Article 4 TDM exception for permissible training-data uses. These are separate legal contexts requiring separate legal treatment.

Training Data Transparency: A Proportionate Middle Ground

Rightsholders may have difficulty determining whether a particular AI model was trained on their works. Without this information, rightsholders cannot verify if a model has respected their opt-out preferences and are at a disadvantage when negotiating voluntary licenses. ITIF believes this concern is legitimate but that the GPAI Code of Practice can address it without much further revision.

The GPAI Code of Practice already requires providers to publish summaries of training datasets, and Article 53(1)(d) of the EU AI Act mandates that all GPAI providers publish a “sufficiently detailed summary” of training content according to a template issued by the AI Office in July 2025.[7] ITIF supports targeted transparency  requirements that enable confirmation of lawful use of copyrighted content and therefore recommends that the Commission clarify and strengthen it in the following respects:

1. Dataset summaries should identify the sources from which it was collected, the time periods covered, and the scale of use.

2. Rightsholders who have a good-faith basis to believe their works were included in a training dataset should have access to a practical mechanism to seek confirmation. This does not require the disclosure of proprietary model weights or training pipelines, but it should enable rightsholders to discover this information with a minimum of friction.

3. Transparency obligations should be minimally burdensome so as not to impose high costs on GPAI providers and chill innovation. GPAI providers include a wide range of capabilities, from large-scale commercial GPAI providers to academic researchers and small developers training specialized models.

Targeted transparency provisions in this way would allow markets and legal processes to function more efficiently; mandatory licensing substitutes a regulatory mechanism for market pricing. We respectfully suggest that the Commission should treat these as independent policy levers.

Performer Impersonation: A Targeted Rights Problem Requiring a Targeted Solution

The call for evidence identifies AI-generated imitations of performers’ voices, likenesses, and personal characteristics as a distinct challenge, raising questions that go beyond copyright protection. ITIF agrees with this framing and urges the Commission to treat it accordingly: as a targeted rights problem requiring a targeted solution, not as a basis for expanding copyright in ways that would inadvertently restrict AI development more broadly.

A. Performer Protection Belongs in Personality Rights Law, Not Copyright

Under traditional copyright doctrine, a performer's voice, likeness, and personal characteristics are not themselves copyrightable works in the EU, the United States, or most major jurisdictions. Copyright protects original creative works (a recording, a film, a performance), not the underlying personal attributes of the individual. These attributes are instead protected, where protected at all, through separate legal regimes: personality rights in EU member states (e.g., droit à l'image in France, allgemeines Persönlichkeitsrecht in Germany), right of publicity and limited trademark theories in the United States, and no codified statutory right at all in the UK currently.[8]

However, this categorical separation is no longer universal. Denmark is considering a law that reclassifies voice and physical appearance as a neighbouring right to copyright, making them potentially transferable and commercially exploitable IP assets rather than inalienable personality rights, which blurs a previously well-established legal distinction.

Notwithstanding the foregoing, ITIF suggests that the appropriate legal framework for protecting performers against unauthorized AI-generated imitations is personality rights, right of publicity, or an equivalent dedicated instrument rather than an expansion of copyright or related rights that would risk narrowing the TDM exception or creating new authorization requirements for AI training on publicly available content. The Commission has an opportunity to adopt a more coherent approach from the outset by designing a harmonised EU instrument that is precisely scoped to unauthorized commercial impersonation.

B. Key Design Principles

If the Commission proceeds with a legislative measure on performer impersonation, ITIF recommends the following principles to ensure the measure is effective without creating unintended barriers to AI innovation:

1. Scope to commercial, deceptive, or harmful uses. The harm arises from unauthorized commercial exploitation, audience deception, and reputational damage, not from the technical act of training a model on content that includes a performer’s voice or image. The measure should target the output and its use, not the training process.

2. Protect all performers, not only celebrities. Performers at every level of public recognition have a legitimate interest in controlling commercial use of their voice and likeness. Remedies should be scaled proportionately to the harm caused.

3. Include explicit safe harbours for legitimate uses. Dubbing, language accessibility tools, archiving, parody, satire, education, and transformative creative expression all serve genuine public interests. Clear safe harbours for these uses are essential to avoid chilling beneficial AI applications and protecting free expression.[9]

4. Confirm that this measure does not reach training data. The Commission should explicitly state that a performer protection instrument does not create new authorization requirements for the use of publicly available recordings in AI training. The TDM exception already governs that use and should remain unaffected.

Competitiveness Implications for the Transatlantic Ai Ecosystem

A. Regulatory Asymmetry Creates Real Structural Disadvantages

The United States maintains a relatively permissive framework for AI training under the fair use doctrine. ITIF acknowledges that the U.S. fair use landscape is not fully settled: while some courts have found AI model training to be transformative use, others have declined to extend fair use protection where AI training was closely tied to producing a directly competing product.[10] The Commission should be cautious about overstating the permissiveness of U.S. law. That said, the overall U.S. framework remains substantially less restrictive than a mandatory licensing regime would be, and EU developers operating under such a regime would bear costs that their U.S. counterparts do not.

China is also closing the AI capability gap rapidly.[11] The EU’s regulatory choices will influence whether European developers can compete at the frontier of AI, or whether the EU cedes that ground to non-European actors. ITIF urges the Commission to weigh competitiveness implications explicitly in its impact assessment — not as a reason to ignore legitimate creator interests, but as a necessary input to a genuinely balanced policy calculus.

B. The Precautionary Approach has Quantifiable Costs

ITIF has consistently argued that the EU’s reliance on precautionary regulation in technology policy imposes significant costs on innovation and economic growth. In the AI training context, a precautionary framework that presumes any use of copyrighted content in training is harmful, absent explicit authorization, would create a uniquely restrictive environment in Europe. The Commission should adopt an evidence-based approach: identify specific, demonstrated harms; target regulatory interventions precisely at those harms; and preserve the TDM exception as the default.

This does not mean dismissing creator concerns. It means calibrating the regulatory response to the nature and magnitude of the harm. Output-side reproduction of copyrighted content is a demonstrated harm with identifiable victims and is addressable through existing legal tools. The harm from training-data use is more diffuse, contested, and difficult to measure—and the Commission’s approach to each should reflect that difference.

The Research Exception and Out-Of-Commerce Works: Pro-Innovation Opportunities

Two further issues raised in the call for evidence are directly relevant to ITIF’s positions and deserve explicit engagement: the harmonisation of the research exception under the InfoSoc Directive, and the availability of out-of-commerce works for AI training under the CDSM Directive.

A. A Harmonised Research Exception Advances Innovation

The call for evidence notes that the optional research exception under the InfoSoc Directive is implemented inconsistently across Member States, creating legal uncertainty for cross-border research collaboration. The Commission is considering clarifying or further harmonising this exception and potentially introducing an EU-level secondary publication right that would allow authors of publicly funded research to make their work openly available independently of the publisher’s version.

ITIF strongly supports both measures. A harmonised, robust research exception helps reduce fragmentation in the legal treatment of research activities across Member States, lowers compliance costs, supports cross-border collaboration, and increases the availability of high-quality scientific content for downstream uses, including AI training. This directly reinforces the objectives of the TDM framework and should be pursued as a complementary measure.

Additionally, research funded by public institutions generates knowledge that benefits from broad dissemination. Authors of publicly funded research commonly lack bargaining power relative to scientific publishers, resulting in rights transfers that restrict access to and reuse of knowledge that the public has already paid to produce. A secondary publication right that enables authors to make their work openly available after an appropriate embargo period, noting any contrary contractual term, corrects this imbalance without disrupting the incentives of commercial publishers to invest in the publication and peer review process. ITIF supports the Commission’s exploration of this option and encourages it to draw on the experience of Member States that have already introduced such rights at the national level.

B. Out-of-Commerce Works Must Remain Available for AI Training

The Commission’s review survey raises a pointed question about whether out-of-commerce works made available under the CDSM Directive should remain available for AI model training. ITIF urges the Commission to confirm clearly that they should.

Out-of-commerce works, i.e., those no longer commercially exploited by rightsholders, are precisely the category of content for which the case for unrestricted TDM access is strongest. These works do not carry a meaningful risk of market substitution (there is no active market to displace), yet they represent an enormously valuable cultural, linguistic, and historical corpus. Restricting their availability for AI training would reduce the quality and cultural breadth of AI systems trained on European content, without corresponding benefit to any identifiable rightsholder. It would also conflict with the public interest objectives that justified making these works accessible in the first place.

ITIF recommends that the Commission should confirm that the Article 3 TDM exception for scientific research and the Article 4 general TDM exception apply to out-of-commerce works made available under the CDSM Directive, and that no new restrictions on this use are warranted.

Online Piracy and Equitable Remuneration

The call for evidence addresses two additional areas: online piracy of live events, and the application of the single equitable remuneration right to sound recordings of third-country nationals.

On online piracy of live events, the Commission is exploring stronger enforcement remedies and cross-border cooperation mechanisms for the protection of sports broadcasts and other live content. ITIF supports legal frameworks authorizing the blocking of piracy websites used to stream live events. ITIF notes that best practices for such legal frameworks include ensuring that courts retain exclusive authority to issue blocking orders; ISPs receive reasonable immunity from liability for complying with blocking orders; ISPs are fairly compensated for the costs incurred in implementing blocking measures; and blocking injunctions are enforceable across all ISPs.[12]

On the single equitable remuneration right and third-country nationals, the Commission is considering introducing a principle of material reciprocity following the Court of Justice’s judgment in Case C-265/19 (RAAP). This is a specialized question in international copyright law on which ITIF does not offer a view. ITIF notes only that any reciprocity mechanism should be designed with appropriate safeguards to avoid unintended effects on the digital music ecosystem or on the licensing relationships that support cross-border distribution of recorded music.

Recommendations

ITIF recommends that the Commission consider structuring the Targeted Initiative around the following principles:

1. Preserve the Article 4 TDM exception as the foundational framework for AI training on lawfully accessible content, including out-of-commerce works made available under the CDSM Directive. The Commission should not reopen, narrow, or add mandatory compensation requirements to the existing exception.

2. Distinguish training from output in all legislative or guidance measures. Any new rules addressing AI output reproduction should be explicitly scoped to that context and should not affect or narrow the TDM exception.

3. Address output-side reproduction harms through targeted enforcement guidance. The Commission should clarify how existing copyright law applies to AI systems that reproduce or substitute for copyrighted content in real-time outputs, and support enforcement of existing rights rather than creating new ones.

4. Adopt technically feasible, accessible opt-out standards. Required protocols should be based on robots.txt and HTTP headers, apply prospectively, and be supplemented by collective opt-out infrastructure to reduce practical barriers for small and independent creators.

5. Strengthen GPAI training data transparency requirements. Dataset summaries should be substantively informative, and a proportionate mechanism should exist for rightsholders to seek confirmation that their works were not used in training.

6. Reject mandatory collective licensing for training data. The Commission should allow voluntary licensing arrangements to develop through market negotiation and should not substitute a regulatory pricing mechanism for market-determined terms.

7. Introduce a targeted performer protection instrument, scoped to unauthorized commercial impersonation, grounded in personality rights rather than copyright, with explicit safe harbours for legitimate uses. Confirm expressly that this instrument does not affect AI training data use.

8. Prioritize voluntary mediation and arbitration for AI licensing disputes. Any dispute resolution framework should be voluntary, technically competent, coordinated with the GPAI Code of Practice, and not used as a basis for mandatory rate-setting.

9. Apply the human authorship standard consistently to AI-generated and AI-assisted works. The Commission should not create new IP rights in purely AI-generated outputs or impose AI-content disclosure as a condition of copyright protection.

10.Harmonise the research exception and introduce an EU-level secondary publication right for publicly funded research. Both measures are pro-innovation, reduce legal fragmentation across Member States, and increase the availability of high-quality content for downstream uses, including AI training.

11.Conduct a rigorous, balanced competitiveness impact assessment. The Commission should explicitly assess how each proposed option affects European AI developers’ competitive position relative to U.S. and Chinese counterparts, alongside its assessment of creator impacts.

Conclusion

The Commission’s call for evidence spans four distinct policy areas—AI and copyright, online piracy, performer remuneration, and research—and reflects the breadth of challenges that digital technology poses to the existing EU copyright framework. ITIF has addressed in these comments the areas most directly relevant to its expertise in AI and innovation policy: the preservation of the TDM training exception; the importance of distinguishing training-data use from output-side reproduction; the need for workable transparency and opt-out mechanisms; the appropriate legal vehicle for addressing performer impersonation; the value of voluntary dispute resolution over mandatory licensing; and the pro-innovation case for harmonising the research exception and introducing a secondary publication right.

Across all of these areas, the common thread in ITIF’s recommendations is proportionality: regulatory interventions should be calibrated to the nature and magnitude of the harm they address, should not restrict beneficial uses of AI in pursuit of harms that existing law already addresses, and should preserve the open information environment that has underpinned both technological innovation and European cultural production. The Commission can advance both creativity and competitiveness—but only if it keeps these goals analytically distinct and designs its legislative tools accordingly.

ITIF thanks the Commission for the opportunity to submit these comments and remains available to provide additional input or technical assistance as the Initiative develops.

Endnotes