EU AI Copyright Laws: Opt-Outs and AI Act Duties (2027)

EU AI copyright law, as of November 4, 2026, rests on two instruments: the text and data mining exceptions in Directive (EU) 2019/790 (the DSM Directive) and the copyright duties for general-purpose AI model providers in Article 53 of the AI Act, Regulation (EU) 2024/1689. Both govern training. For AI output, EU law looks for a human author, and the Court of Justice has not yet ruled in a generative AI case.

Status as of November 4, 2026: General-purpose AI duties have applied since August 2, 2025, and the Commission's fining power under Article 101 has applied since August 2, 2026. Models placed on the market before August 2, 2025 have until August 2, 2027. In Like Company v Google Ireland (C-250/25), the Advocate General delivered a non-binding Opinion on November 3, 2026, and judgment is still to come. The Commission's review of the DSM Directive is under way.

Scope: This article covers EU-level law and selected German rulings on AI output, training and provider duties. It is general information, not legal advice. For a comparative overview, see how AI and copyright differ worldwide.

Can AI-Generated Works Be Copyrighted in the EU?

The European Parliament's March 10, 2026 resolution states that EU copyright law "remains grounded in the principles of human authorship." It describes the Court of Justice's settled case-law as setting two cumulative conditions for a work: original subject matter that "reflects the author's own intellectual creation," expressed with sufficient precision and objectivity.

Parliament went on to insist that content fully generated by AI that does not meet those criteria "should remain ineligible for copyright protection," and it asked that the public domain status of such output "be clearly determined." That wording matters. A resolution is a political position and not a ruling, and asking for the status to be determined suggests the question is not closed.

The Court of Justice's own press release calls Like Company "the first generative AI case to reach the Court of Justice." That case is about press publishers' rights in chatbot answers, not about whether AI output is a protected work. Parliament says content is classified as AI-generated or human-created depending on whether the output meets the established criteria for copyright protection.

Text and Data Mining: Articles 3 and 4 of the DSM Directive

The Commission describes the DSM Directive as introducing "two new mandatory exceptions or limitations for text and data mining." AI Act recital 105 explains why they matter here: training needs large amounts of content, and using protected content generally requires authorisation unless an exception applies.

The Article 4 exception applies on two conditions, as the Code of Practice copyright chapter sets them out: lawful access, and no express reservation of the use by the rights holder. The reservation must be made "in an appropriate manner, such as machine-readable means in the case of content made publicly available online," in the Commission's words.

Recital 105 says rights holders may opt out unless the mining is done for scientific research. It adds that where the opt-out has been expressly reserved in an appropriate manner, providers of general-purpose AI models need authorisation from the rights holder.

Article 3 of the Directive is limited by who benefits as well as by purpose. It covers research organisations and cultural heritage institutions mining, for the purposes of scientific research, material they have lawful access to. Article 2 defines a research organisation by a not-for-profit, reinvestment or public-interest mission, and bars preferential access to results for an undertaking with decisive influence over it. A research purpose alone does not bring a developer within Article 3.

What counts as a valid reservation is unsettled. The Commission says the meaning and scope of Article 4(3) "will ultimately be a matter for the Court of Justice." The Commission opened a consultation on opt-out protocols on December 1, 2025, closed it on January 23, 2026, and said it will publish a list of generally agreed machine-readable solutions, reviewed at least every two years.

AI Act Article 53: Copyright Duties for General-Purpose AI Providers

Under Article 53(1)(c), providers must "put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3)" of the DSM Directive.

Under Article 53(1)(d), they must "draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office."

Open-source models. Article 53(2) lifts only the documentation duties in points (a) and (b) for models released under a free and open-source licence. It does not lift points (c) and (d). The open-source relief also does not apply to general-purpose AI models with systemic risks.

Providers outside the EU. Article 2(1)(a) covers providers placing general-purpose AI models on the market "irrespective of whether those providers are established or located within the Union." Recital 106 says a provider placing a model on the Union market should comply with the copyright-policy obligation "regardless of the jurisdiction in which the copyright-relevant acts underpinning the training" take place.

That recital concerns the AI Act duty. Whether EU copyright law itself reaches training done abroad is a different question. Parliament's resolution urges that reading, which is a position and not a ruling.

Fines. Article 101(1) lets the Commission fine general-purpose AI providers up to 3% of annual total worldwide turnover in the preceding financial year or EUR 15,000,000, whichever is higher, for intentional or negligent infringement. Article 113(b) carves Article 101 out of the August 2, 2025 start date, so it applies from the general date of August 2, 2026. The Commission describes the AI Office as having "effective enforcement powers as from August 2026."

Older models. Under Article 111(3), providers of models placed on the market before August 2, 2025 "shall take the necessary steps" to comply by August 2, 2027.

The Training Data Summary Template

The Commission published the template and an explanatory notice on July 24, 2025. Recital 107 says the summary should be "generally comprehensive in its scope instead of technically detailed," with due regard for trade secrets, for example by listing the main data collections and explaining other sources in narrative form.

According to the Commission, summaries should describe public and private datasets, list all large publicly available datasets, and give information on data scraped from online sources. That covers crawler names, collection period, the type of content crawled, and the top 10% of scraped domains. For small and medium-sized enterprises, the figure is the top 5% or 1,000 domains, whichever is lower.

The Commission says the AI Office will monitor whether summaries conform to the template. It also recommends, on a voluntary basis, that providers let rights holders ask for more information about their works on specific internet domains.

The GPAI Code of Practice Copyright Chapter

The General-Purpose AI Code of Practice was published on July 10, 2025. It is a voluntary compliance tool, and its Copyright Chapter says adherence "does not constitute compliance with Union law on copyright and related rights." It also says the Code is not conclusive evidence of compliance with the AI Act.

The Commission's page listed 22 signatories when read on November 4, 2026, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. It says xAI signed only the Safety and Security chapter and must show compliance with the copyright obligations through other adequate means. The AI Office supervises Article 53 compliance for all providers, signatories or not.

Signatories commit to the following for models they place on the Union market:

• Keep a single-document copyright policy, assign responsibility for it, and keep it up to date.
• Crawl only lawfully accessible content: do not circumvent effective technological measures such as paywalls, and exclude websites recognised by EU or EEA courts or authorities as persistently infringing on a commercial scale.
• Follow robots.txt as specified in IETF RFC 9309, and comply with other appropriate machine-readable reservation protocols.
• Publish information about their crawlers and how they identify reservations.
• Use appropriate and proportionate technical safeguards against outputs that reproduce training content in an infringing manner, and prohibit infringing uses in their terms.
• Name a point of contact and let rights holders file complaints electronically.

The Code says these commitments do not affect the right to reserve rights "in any appropriate manner," so it does not treat robots.txt as the only way to opt out.

Like Company v Google Ireland (C-250/25)

Like Company, a Hungarian web portal operator, sued Google before the Budapest Regional Court, claiming protection under the press publishers' right in Article 15(1) of the DSM Directive. The Hungarian court sent four questions to the Court of Justice. Two concern training and two concern chatbot answers.

On November 3, 2026, First Advocate General Szpunar advised the Court to declare the two training questions inadmissible as hypothetical. The press release gives two reasons: there is no evidence Like Company's articles were used to train Gemini, and training is not an "online use" of press publications. For completeness, he suggested that using protected material to train a large language model is, "as a rule, freely permitted under the text and data mining exception."

On chatbot answers, he advised that where a chatbot retrieves a publication from a press publisher's website, reuses it to generate an answer and gives that answer to the user, the press publisher's right "may be infringed in certain circumstances." The right protects the form of expression and not the facts reported. Under his approach, the reproduction right turns on two factors:

• Whether the chatbot accessed the publication unlawfully, in particular by circumventing technical measures such as a paywall.
• Whether the answer merely reports facts, or reuses the form of expression recognisably, to the point that the user need not read the original.

He added that making available to the public is, as a rule, infringed where that form of expression is systematically recognisable in answers to related queries. A provider would be liable where such acts are inherent to its business model, or where it failed to put appropriate measures in place to counter them credibly and effectively.

An Opinion is not binding on the Court. The judges have begun deliberating and judgment will come at a later date. These points come from the Court's press release, which is an unofficial document for media use.

German Courts: LAION, GEMA v OpenAI and GEMA v Suno

The national rulings below apply German law. None of them comes from the Court of Justice.

Hamburg. A non-profit downloaded a photo to compare it with its description while building a free dataset of image-text pairs usable for AI training. The court held the agency's reservation was attributable to the photographer, but it did not have the required machine-readable form at the time of download. It also held the use was justified as scientific research.

The Federal Court of Justice's May 20, 2026 notice set a hearing for September 3, 2026. The photographers' association BFF, which backs the photographer's side, reported on September 9, 2026 that nothing was decided that day and that the senate signalled it will refer questions to the Court of Justice. The Federal Court of Justice's own notice sets a decision date of December 17, 2026. We read no court statement confirming the reported signal.

Munich. In GEMA v OpenAI, the court found nine songwriters' lyrics reproducible in GPT-4 and 4o, and held that storing them in model parameters is a reproduction. It accepted that mining exceptions can apply to models in principle, but said training that memorises whole works is not text and data mining. It held the defendants, not users, responsible for outputs.

In GEMA v Suno, the same chamber found memorisation in models stored on servers in Germany, held it was not covered by the German mining exception, and held the defendant, not users, responsible for outputs generated from simple prompts. The release says it also rejected a US fair use defense for copies made in the United States.

How the EU Differs from the United States

The EU has a written reservation mechanism for mining and a regulator-enforced copyright policy and disclosure duty for general-purpose AI providers. US disputes over training have turned on fair use in the courts. See AI copyright law in the United States for those rulings.

What Changes in EU AI Copyright Law in 2027

Two items below are fixed legal dates: August 2, 2027 and, in the last item, December 2, 2026. The rest are expected or pending steps.

• August 2, 2027: General-purpose AI models placed on the market before August 2, 2025 must comply (AI Act Art. 111(3)).
• First quarter of 2027 (expected): A May 8, 2026 Council note says the Commission is expected to present an AI Strategy for Cultural and Creative Sectors.
• C-250/25: Judgment is pending as of the Court's November 3, 2026 release, with no date stated.
• DSM Directive review: A call for evidence ran to June 25, 2026. A consultation on technology and copyright opened on September 29, 2026 and was extended to November 10, 2026. The Commission says it will examine the mining exceptions and explore whether adjustments or additional measures are needed. We read no legislative proposal as of November 4, 2026.
• Opt-out registry: A study published July 13, 2026 found an EU registry could be "a useful complementary instrument." The Commission says it will consider next steps.
• Before 2027: The Federal Court of Justice decision date of December 17, 2026 could send the LAION questions to the Court of Justice. Separately, AI systems generating synthetic content placed on the market before August 2, 2026 must meet the Article 50(2) marking duty by December 2, 2026, under the amended Article 111(4).

Change Log

• November 4, 2026: Rewritten against official sources. Added the Advocate General's Opinion in C-250/25, the Code of Practice copyright chapter, the training summary template, recital 106, the Hamburg and Munich rulings and the DSM Directive review. Corrected the description of who is covered outside the EU, added the systemic-risk limit on open-source relief, and added that fines are Commission-imposed and began August 2, 2026. Removed the sections on case-law, database rights and software that this rewrite did not re-verify against their source texts.
• September 10, 2026: Corrected the grandfathering date for general-purpose AI models: Article 111(3) covers models placed on the market before August 2, 2025.
• September 9, 2026: AI Act dates updated for the July 2026 Digital Omnibus (Regulation (EU) 2026/1744).

Related Articles

• How AI and copyright differ worldwide
• AI copyright law in the United States
• United Kingdom AI copyright laws
• Japan AI copyright laws

Frequently Asked Questions

Does the EU grant copyright to AI-generated art or text?

The European Parliament said in March 2026 that EU copyright law is grounded in human authorship and that fully AI-generated content not meeting the established criteria should stay ineligible. That is a resolution, not a ruling. As of November 4, 2026, the Court of Justice has not ruled on the question in a generative AI case.

Can EU rights holders stop AI companies from training on their content?

Rights holders can expressly reserve their works from general text and data mining under Article 4 of Directive (EU) 2019/790 in an appropriate manner, such as machine-readable means for content published online. AI Act recital 105 says providers then need authorisation. The opt-out does not reach mining under Article 3, which covers research organisations and cultural heritage institutions acting for scientific research, and the scope of a valid reservation is still for the Court of Justice to settle.

Is robots.txt a valid opt-out under EU law?

Code of Practice signatories commit to following robots.txt, but the Code says rights holders may reserve rights in any appropriate manner. The Commission says the meaning of Article 4(3) will ultimately be for the Court of Justice. A German appeal court held that the reservation in the case before it was not machine-readable.

When do EU AI Act copyright obligations apply?

General-purpose AI obligations have applied since August 2, 2025. Article 101 fines have applied since August 2, 2026. Models placed on the market before August 2, 2025 have until August 2, 2027.

Do open-source AI models have to follow the EU copyright rules?

Yes. Article 53(2) lifts only the documentation duties in Article 53(1)(a) and (b) for free and open-source models. The copyright policy and training summary duties in points (c) and (d) still apply.

Does the EU AI Act apply to AI companies outside the EU?

Yes, for providers placing general-purpose AI models on the EU market, wherever they are established. Recital 106 says the copyright-policy duty applies regardless of where the training took place. Whether EU copyright law itself reaches training done abroad is a separate, debated question.

What did the Advocate General say in Like Company v Google?

On November 3, 2026, Advocate General Szpunar advised that a chatbot that retrieves press publications in real time and reuses them to generate answers may infringe press publishers' rights in certain circumstances, and that the two training questions are inadmissible. For completeness, he suggested training is, as a rule, permitted under the mining exception. The Opinion does not bind the Court, which has not yet ruled.

Is memorising songs in an AI model copyright infringement in Germany?

Munich Regional Court I held so in GEMA v OpenAI on November 11, 2025 and GEMA v Suno on July 31, 2026. Each press release says the judgment is not final.