EU copyright, TDM opt-outs and the AI Act: what applies to legal training data

The EU text and data mining (TDM) rules apply to legal training data, but for most official legal texts they have little to bite on. Directive (EU) 2019/790 (the DSM Directive) lets anyone mine lawfully accessible works unless the rightholder has “expressly reserved” that use (Article 4(3)), and Article 53(1)(c) of the AI Act, Regulation (EU) 2024/1689, requires every general-purpose AI model provider to have a policy that identifies and complies with those reservations, in force since 2 August 2025. An opt-out reserves a copyright or database right, though, and statutes and court decisions are excluded from copyright in most member states, while EUR-Lex content is licensed for reuse under its legal notice. What still needs care is editorial content, third-party material embedded in legal texts, publisher terms, and personal data, which copyright exceptions do not touch.

What do the TDM exceptions say?

The DSM Directive defines text and data mining as “any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations” (Article 2(2)). It creates two exceptions:

Provision	Who benefits	Opt-out possible?	Retention
Article 3	Research organisations and cultural heritage institutions, for scientific research	No	Stored securely, retained for research and verification
Article 4	Anyone, including commercial AI developers	Yes, under Article 4(3)	“For as long as is necessary for the purposes of text and data mining”

Article 4(3) is the opt-out: the exception applies “on condition that the use of works and other subject matter referred to in that paragraph has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. Recital 18 adds that, for content made publicly available online, a reservation should only be considered appropriate if it uses machine-readable means, “including metadata and terms and conditions of a website or a service”.

Whether training a generative model is “text and data mining” has not been decided by the Court of Justice. In Case C-250/25 Like Company v Google Ireland, Advocate General Szpunar’s opinion of 3 November 2026 proposed that the questions on LLM training be declared inadmissible as hypothetical, while adding, for completeness, that using works to train an LLM is as a rule permitted under the TDM exception. The opinion is not binding and the judgment may not decide the point. In Germany, the Hamburg courts applied the TDM exceptions to building the LAION image-text dataset (Landgericht Hamburg, 27 September 2024, 310 O 227/23; on appeal, Oberlandesgericht Hamburg, 10 December 2025, 5 U 104/24, which dismissed the appeal and admitted a further appeal on points of law to the Federal Court of Justice).

What does the AI Act add?

The AI Act does not change copyright law. It adds a compliance duty for providers of general-purpose AI models. Article 53(1)(c) requires them to “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 Directive (EU) 2019/790”. Recital 106 makes the duty extraterritorial in effect: any provider placing a model on the EU market should comply “regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place”.

Article 53(1)(d) adds the public training-content summary, whose Section 3.1 asks what the provider did to respect reservations, including those honoured by third parties from which datasets were obtained. That template is covered in AI Act Article 53 training-data summary: a template for legal corpora.

Date	What happens
10 July 2025	General-Purpose AI Code of Practice published, with chapters on transparency, copyright, and safety and security
24 July 2025	Training-content summary template published
1 August 2025	Commission and AI Board confirm the Code is an adequate voluntary tool to demonstrate compliance
2 August 2025	GPAI obligations, including Article 53, apply (Article 113)
1 December 2025	Commission opens a consultation on protocols for reserving rights from TDM
2 August 2026	The AI Office’s enforcement powers, including fines, apply
2 August 2027	Deadline for models placed on the market before 2 August 2025

What does the Code of Practice ask for on copyright?

The copyright chapter of the Code of Practice turns Article 53(1)(c) into concrete measures for signatories. They commit to:

• employ web crawlers that read and follow robots.txt as specified in RFC 9309 (the Robots Exclusion Protocol, September 2022);
• identify and comply with other appropriate machine-readable protocols for Article 4(3) reservations;
• not circumvent effective technological measures such as paywalls;
• exclude websites recognised by EU or EEA courts or authorities as persistently infringing copyright on a commercial scale;
• designate a point of contact and accept complaints from rightholders.

Machine-readable opt-out signals in use include robots.txt rules for AI crawlers and the TDM Reservation Protocol (TDMRep), a W3C Community Group report of 10 May 2024 that uses a tdm-reservation HTTP header or a /.well-known/tdmrep.json file. The Code is voluntary; a provider that does not sign must show compliance in another way.

Why do opt-outs matter less for official legal texts?

An Article 4(3) reservation reserves a right the rightholder has. If no copyright or database right exists in a text, there is nothing to reserve, and the TDM exception is not needed in the first place.

International law leaves official texts to national law. Article 2(4) of the Berne Convention says: “It shall be a matter for legislation in the countries of the Union to determine the protection to be granted to official texts of a legislative, administrative and legal nature, and to official translations of such texts.” Most European countries exclude them:

Country	Rule	Effect
Germany	§ 5(1) UrhG: statutes, ordinances, official decrees and notices, decisions and official headnotes “genießen keinen urheberrechtlichen Schutz”	No copyright in federal or Land law or court decisions
Netherlands	Auteurswet art. 11: no copyright in laws, decrees and ordinances issued by public authorities; Databankenwet art. 8 on databases of laws and judgments	No copyright or database right for statutes and judgments
Poland	Art. 4 Prawo autorskie	Official acts and judgments excluded
Denmark, Sweden, Croatia, Lithuania	Equivalent statutory exclusions	Official texts outside copyright
EU	No exclusion: EUR-Lex content is “© European Union”	Reuse is licensed by the EUR-Lex legal notice instead
UK	Legislation is Crown copyright	Licensed under the Open Government Licence v3.0

So for EU law the question is the scope of an express licence, not an exception, and for most national statutes the text is free of copyright altogether. The country-by-country terms are in National legislation datasets in Europe.

What still needs care?

Copyright-free text does not mean everything around it is free. Five things still matter:

1. Editorial and third-party material. Headnotes written by private publishers, commentary, summaries and translations may be protected. EUR-Lex excludes third-party intellectual property, such as standards referenced in acts and images, from its reuse terms, and CURIA warns that judgments sometimes reproduce protected works such as trade-mark images.
2. Database rights. The sui generis right in Directive 96/9/EC protects “a substantial investment in either the obtaining, verification or presentation of the contents” of a database for 15 years. The Directive has no general exclusion for official databases, so whether a public legal database attracts the right depends on national law; the Netherlands excludes databases of laws and judgments expressly.
3. Contract and access terms. A publisher’s terms of use or API conditions bind you even where copyright does not. Some legal sources restrict uses outright: UK court judgments on Find Case Law need a separate licence for computational analysis, including AI training.
4. Robots.txt as a signal. Several official portals block AI crawlers. Where no right exists to reserve, a block is not an Article 4(3) opt-out, but it is a clear statement of the publisher’s wishes and a factor regulators weigh when assessing data collection, so use the documented API or bulk channel the publisher offers and record why. The differences between crawled and curated data are in Common Crawl vs curated legal data.
5. Personal data. Court decisions name or pseudonymise people. Copyright exceptions do not affect data-protection law, and the GDPR applies to training on them in full.

What should a provider document?

For a legal corpus in the copyright policy and training-content summary, record per source:

Item	Example for a legal corpus
Legal basis for reuse	Statutory exclusion from copyright, an open licence (OGL, Licence Ouverte, CC BY) or the EUR-Lex legal notice
Access channel	Official API, bulk download or licensed supplier, not a generic crawl
Reservation signals	Robots.txt and TDMRep status of the source, and how they were handled
Exclusions	Sources not used because their terms exclude AI training
Third-party material	Referenced standards, images and other excluded content
Personal data	Whether court decisions were included, and how they were handled

A licensed corpus that records this for you shortens the work: you report it under Section 2.2 of the summary and cite the supplier’s per-source record in Section 3.1.

Key facts

• The DSM Directive’s Article 4 TDM exception applies unless the rightholder expressly reserves it, in a machine-readable way for content online (Article 4(3), recital 18).
• AI Act Article 53(1)(c) requires GPAI providers to identify and honour those reservations; it has applied since 2 August 2025, with AI Office enforcement from 2 August 2026.
• The GPAI Code of Practice (10 July 2025) commits signatories to follow robots.txt under RFC 9309 and other machine-readable opt-out protocols.
• Official legislative and judicial texts are outside copyright in Germany, the Netherlands, Poland and many other countries; EU texts are licensed for reuse under the EUR-Lex legal notice.
• The CJEU has not ruled on whether LLM training is text and data mining; the AG opinion in C-250/25 (3 November 2026) is not binding.

Frequently asked questions

Can a government portal opt out of AI training under Article 4(3)?

Only for rights it holds. Where statutes and judgments are excluded from copyright, and no database right applies, there is no right to reserve. The portal’s terms of use and robots.txt still express its wishes and may bind you contractually.

Does the AI Act copyright policy apply if I train outside the EU?

Yes, if you place the general-purpose AI model on the EU market. Recital 106 says providers should comply regardless of where the copyright-relevant acts of training take place.

Is EUR-Lex text covered by the TDM exception?

You do not need the exception for most EUR-Lex content, because the EUR-Lex legal notice licenses reuse, including commercial reuse, with acknowledgement of the source. Third-party material on EUR-Lex is excluded from that licence.

Does the TDM exception cover personal data in court decisions?

No. Copyright exceptions do not affect data-protection law. Training on court decisions that identify people is personal-data processing and needs a lawful basis under the GDPR.

Where Arieta Data fits

Arieta Data records the legal basis, licence, attribution and access channel for every source, and every record carries its source and licence, so the copyright policy and training-content summary have something concrete to cite. It ships as the corpus, the history and the graph. Snapshot 2026-11 holds 419,355 documents from 16 jurisdictions, in 24 languages (EU documents are almost all in English; the national lanes cover 14 national languages): EU legislation and CJEU case law reused under the EUR-Lex legal notice and the CURIA copyright notice, and national legislation from 15 countries under their open licences or statutory exclusions. UK case law is excluded because the Open Justice Licence does not cover computational analysis. National case law is included, delivered with personal data masked. The per-source record is on the coverage page and the method on the provenance page.

Request a sample and the per-source licence record. We reply within one business day, and a custom cut by jurisdiction, language, date range or document type is ready within 5⁠–⁠6 business days.

Not legal advice. This guide summarises public sources as of 13 November 2026; the law on AI training and copyright is unsettled, so take your own legal advice.