2026-03-16 - Copyright and Generative AI: The European Parliament's Resolution Between Principles and Operational Gaps
On 10 March 2026, the European Parliament adopted in plenary session in Strasbourg Resolution
on copyright and generative artificial intelligence — opportunities and challenges. The procedure was initiated in 2025 (2025/2058(INI)), and the final document is the outcome of work by the Committee on Legal Affairs, with rapporteur Axel Voss. That represents the first formal position of the new European Parliament on one of the most legally complex issues of the current EU regulatory cycle: the relationship between copyright-protected works and the training processes of generative AI systems.
The Resolution introduces no new binding rules, but clarifies with uncommon precision three points of friction between copyright and generative AI: the operational failure of the DSM Directive’s opt-out mechanism, the extension of transparency obligations beyond training to inference and RAG, and the need to apply Union law to AI providers operating in the European market regardless of where model training takes place. These three axes frame the analysis that follows.
Context: An Expected and Necessary Document
The Resolution emerges at a moment of peak regulatory pressure on AI. The AI Act (Regulation EU 2024/1689) has entered into force, but does not address the specific intellectual property questions raised by model training. The Code of Practice for General-Purpose AI (GPAI Code of Practice), published by the European Commission on 10 July 2025 and in force since 2 August 2025, addresses the topic in its dedicated copyright chapter, but it is entirely voluntary, and not all AI providers have subscribed
to it.
It is in this context that the European Parliament intervenes, identifying three fundamental unresolved legal questions (recital M): the lawfulness of using protected works in training datasets; the legal status of AI-generated content; and the conditions for ensuring
transparency, consent, and fair remuneration for creators.
Three central questions. No binding answers in the Resolution’s text.
The DSM Directive and the Structural Failure of Opt-Out
The first critical issue concerns the
Copyright in the Digital Single Market Directive(2019/790)
, in particular Article 4, which introduces a text and data mining (TDM) exception for commercial purposes, accompanied by an opt-out mechanism: rightholders may expressly reserve their content, excluding it from TDM activities.
The Resolution openly acknowledges that this system does not work. Existing opt-out mechanisms are often impractical, fail to cover all relevant actions connected to text and data extraction, and lack the transparency necessary for effective implementation (recital K). The 2019 norm, conceived for a pre-generative digital context, proves structurally inadequate when applied to the massive training of large language models.
That is, on closer examination, the most important implicit acknowledgement in the entire document: the DSM Directive is not simply being poorly applied — it is structurally inadequate for generative AI. The TDM exception in Article 4 was designed for research and analysis activities on circumscribed datasets; its transposition to the training cycle of generative models, which ingests billions of parameters from heterogeneous and global sources, renders it a blunt instrument. That is not a question of implementation: it is a question of regulatory design.
The proposed solution — standardised, machine-readable tools managed and catalogued by EUIPO (recital X) — is conceptually sound, but remains aspirational: the Resolution is not a legislative act and introduces no binding legal obligations. The ball is back in the Commission’s court.
Extended Transparency: From Training to Inference and RAG
One of the most significant — and least commented upon — passages from a technical and legal standpoint concerns the extension of transparency obligations. Recital Y clarifies that the obligation to disclose copyright-protected content used does not apply only to the model training phase, but extends — mutatis mutandis — to any subsequent use, including for inference and Retrieval-Augmented Generation (RAG).
That is a statement of considerable reach, one that deserves to be stated in all its implications. Inference and RAG occur continuously and in real time: every response generated by an AI system drawing on third-party content could fall within the scope of transparency obligations. The Resolution specifies that crawlers must identify themselves to web operators and that AI providers must maintain detailed logs of automated data
collection activities. It also introduces an operational presumption: content is deemed to have been subject to automated data collection if AI systems process user requests. And it establishes a principle of significant systemic importance: the mere information about third-party content used by AI providers does not constitute a trade secret under Union law (recital Y).
If this framework were translated into binding norms, the transparency obligation would no longer concern only the training dataset — already difficult to reconstruct ex post — but also the informational architecture of systems integrating external retrieval, opening entirely new technical and legal questions: who bears responsibility for transparency in the provider-deployer chain? With what granularity must logs be maintained? What rights arise for a rightholder who identifies their work in a RAG output? These are questions the Resolution poses without answering, but which delineate the perimeter of the next legislative intervention.
Extraterritorial Application: The EU Market as Leverage
Recital P addresses a structural problem with copyright norms: their application is territorial, while AI model training occurs in jurisdictions that are often different from those in which the content was created or the products are marketed.
The Resolution proposes a principle of universal application: EU copyright rules must apply to all AI providers operating in the European market, regardless of their place of establishment and the jurisdiction in which the copyright-relevant actions were performed. The same logic extends to deployers, not only to model providers. A principle coherent with the approach already taken by the GDPR and the AI Act, but which — once again — the Resolution enunciates without equipping it with a specific enforcement mechanism.
It is worth noting that the conclusion of new licensing agreements cannot be interpreted as a retroactive regularisation of prior unauthorised uses (recital P). An important clarification in a context where some providers are concluding commercial agreements with publishers and rightholders that risk being read erroneously as settlements for past infringements.
The GPAI Code of Practice: A Regulatory Architecture Built on Self-Discipline
The Resolution explicitly acknowledges the most visible limitation of the current system: the Code of Practice for General-Purpose AI is entirely voluntary and not all providers have subscribed to it, particularly its copyright chapter (recital R).
That generates a short-circuit with the AI Act: Article 53 of the Regulation requires GPAI model providers to comply with copyright law and to publish a summary of the content used for training, but partly defers to the Code of Practice as an implementation tool. If the Code remains voluntary and only partially applied, the AI Act’s effectiveness on this point remains compromised.
A regulatory architecture that relies on self-discipline by the regulated parties — in a sector that has already demonstrated, as recital L acknowledges, widespread tendencies towards systematic copyright infringement through unauthorised collection, disregard for rights reservations, and use of infringing sources — does not provide adequate guarantees. That is a problem of institutional design, not of provider goodwill.
The Unresolved Tension: Competitiveness or Fundamental Rights?
The Resolution attempts to reconcile two objectives that, in practice, frequently come into conflict. On the one hand, the “AI Continent” action plan and the EU Competitiveness Compass (COM(2025)0030) aim to remove barriers to AI development. On the other hand, the European cultural and creative sector — representing approximately 6.9% of EU GDP and employing around 8 million people (recital C) — requires effective protection.
The Parliament seeks to reconcile these objectives by noting that competitiveness in generative AI requires access to high-quality content, which, in turn, presupposes a fairly remunerated creative sector (recital B). That is a sound argument in its logic, but it risks remaining abstract in the absence of binding rules governing the remuneration of creators whose content is used in training. Competitiveness and the protection of fundamental rights are not mutually exclusive, but neither do they reconcile themselves spontaneously: they require explicit normative choices.
The Jurisprudence to Come: The Risk of Judge-Made Law
Recital T flags an element deserving strategic attention: case law on the matter is still sparse, but a significant number of judicial decisions are expected in the near future, following legal actions initiated by various right-holders in the EU and around the world against certain AI providers.
That is a realistic forecast. But it conceals a risk the Resolution does not make explicit: in the absence of targeted and timely legislative intervention, copyright in the generative AI ecosystem risks being defined more by jurisprudence and international litigation than by a conscious normative choice of the Union. A body of law built case by case, with uncertain outcomes, fragmented by jurisdiction, and potentially incoherent with the principles enunciated in the Resolution itself.
Concluding Observations
Resolution P10_TA(2026)0066 is a politically and legally significant document. It identifies critical issues with precision, rigorously analyzes them, and formulates clear recommendations to the Commission. Its analytical strength is genuine, particularly in its analysis of the structural failure of the DSM opt-out, the extension of transparency to inference and RAG, and the inadequacy of a system built on voluntary provider self-discipline.
But it remains a Resolution: a non-legislative act, devoid of direct binding effect. The real test will be the Commission’s response. The issues raised — from the restructuring of the opt-out mechanism to the regulation of transparency for inference and RAG, from fair remuneration for creators to the extraterritorial application of EU norms — require specific and binding legislative interventions.
In the absence of targeted legislative action, the concrete risk is that copyright in the generative AI ecosystem will be shaped more by jurisprudence and international litigation than by the Union’s deliberate normative choice. Copyright in the age of generative AI is a fundamental right under pressure: the Resolution acknowledges this with clarity. Now it is for the institutions to protect it with instruments equal to the challenge.
Primary sources:
EP Resolution P10_TA(2026)0066, 10 March 2026
Report A10-0019/2026, Committee on Legal Affairs, 25 February 2026
Directive (EU) 2019/790 — Copyright in the Digital Single Market
Regulation (EU) 2024/1689 — AI Act
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.