The EU text and data mining exception, Article 4 of Directive (EU) 2019/790, lets anyone with lawful access copy works to mine them for any purpose, including commercial AI training, unless the rightholder has reserved that use, for online content in a machine-readable way. The exception sits beside the research exception of Article 3, and a reservation under Article 4(3) is the rightholder's switch that takes content out of it.
The Directive (CELEX 32019L0790) entered into force on 2019-06-06, and member states had to transpose it by 2021-06-07. This page covers the conditions of Article 4, the reservation, whether AI training falls within it, the contract override and the rulings still open as of 2026-11-10. The reservation is what turns mining into a question of AI training data licensing.
What Text and Data Mining Means in EU Law (Article 2(2))
Article 2(2) defines text and data mining as any automated analytical technique aimed at analysing text and data in digital form to generate information, including but not limited to patterns, trends and correlations. The definition is neutral about technique and names no field, so it can reach statistics, search indexing and machine learning alike.
Whether model training is such a technique is the question in case C-250/25, and no judgment answers it yet. The Court's press release on the Advocate General's opinion of 2026-11-03 says training is permitted "as a rule" under the exception, but that statement is obiter and not binding. The training section below sets out what is and is not decided.
The Conditions Article 4 Sets for Commercial Mining
Article 4 applies only when five conditions hold together, and each row names the article that sets it. The table is as of 2026-11-10 and cites Directive (EU) 2019/790.
| Condition | What it requires | Article | Source |
|---|---|---|---|
| Lawful access | the works or other subject matter must be lawfully accessible (no further definition is given in the article text) | Art. 4(1) | eur-lex.europa.eu |
| Acts covered | reproductions and extractions for text and data mining, across copyright (Directive 2001/29/EC Art. 2), database rights (Directive 96/9/EC Arts. 5(a), 7(1)), software (Directive 2009/24/EC Art. 4(1)(a)-(b)) and the press publishers' right (Art. 15(1) of the DSM Directive) | Art. 4(1) | eur-lex.europa.eu |
| Retention | copies may be kept as long as necessary for the mining | Art. 4(2) | eur-lex.europa.eu |
| No reservation | the use is not expressly reserved in an appropriate manner, such as machine-readable means for content made publicly available online | Art. 4(3) | eur-lex.europa.eu |
| No contrary contract | Art. 7(1) protects Arts. 3, 5 and 6 only, so a contract can exclude Art. 4 | Art. 7(1) | eur-lex.europa.eu |
Works covered are those protected by national law on or after 2021-06-07 (Art. 26(1)). A mining act that fails one condition falls back on the rightholder's exclusive rights, so a license is needed.
Research Mining Under Article 3 Versus Commercial Mining Under Article 4
Article 3 is a separate, mandatory exception for research organisations and cultural heritage institutions mining for scientific research, and unlike Article 4 it has no opt-out and cannot be overridden by contract. Article 7(1) makes contract terms contrary to Article 3 unenforceable, which is why a rightholder cannot reserve research mining.
The two articles compare as follows, as of 2026-11-10, from CELEX 32019L0790.
| Point | Article 3 | Article 4 |
|---|---|---|
| Who may mine | research organisations and cultural heritage institutions | anyone with lawful access |
| Purpose | scientific research | any purpose, including commercial |
| Opt-out by rightholder | none | reservation under Art. 4(3) |
| Contract override | unenforceable (Art. 7(1)) | possible (not listed in Art. 7(1)) |
A commercial AI developer relies on Article 4, so a reservation reaches it. National research provisions add detail without changing this: Germany's § 60d UrhG, Finland's 13 b § (research copies not overridable by contract) and Sweden's 15 b § (contrary contract terms void).
How a Rightholder Reserves Content Under Article 4(3)
A reservation takes content out of Article 4 when the rightholder expresses it "in an appropriate manner, such as machine-readable means in the case of content made publicly available online". The reservation must be express, machine-readable means are the Directive's example for online content, and the form for offline content is left open. Member states word this differently, as these three patterns show (statute citations as of 2026-11-09):
- Machine-readable form required online: Germany, § 44b(3) UrhG; Austria, § 42h(6) UrhG; Poland, art. 26³ of the copyright act, with metadata (in force 2024-09-20).
- Website or service terms count alongside machine-readable means: Croatia, art. 188(4)-(5); Ireland, s.53B(3) of the Copyright and Related Rights Act.
- An appropriate manner without machine-readable wording: Sweden, 15 a § of the Copyright Act ("på lämpligt sätt").
The full picture for all 27 member states sits on the parent page, linked in the border section below.
In practice a reservation is read by crawlers. TDMRep is a W3C Community Group Final Report of 2024-05-10, not a W3C Standard, and RFC 9309 (robots.txt) is not legally binding by itself. Both are covered with the AI crawlers and how to allow, block or charge them.
Natural-Language Reservations Before the German Courts
Whether a reservation written in plain words on a website, rather than in code, takes content out of the exception is before Germany's Federal Court of Justice in Kneschke v. LAION (I ZR 281/25), with a pronouncement scheduled for December 17, 2026. The case concerns the reproduction of a photograph to build the LAION-5B dataset, with defenses under §§ 44b and 60d UrhG. The Regional Court of Hamburg gave judgment in 310 O 227/23 on 2024-09-27. On 2025-12-10 the Hanseatic Higher Regional Court (OLG Hamburg, 5 U 104/24) dismissed the photographer's appeal: under § 44b(3) UrhG a reservation is effective only in machine-readable form, and the claimant had not shown that the agency's natural-language reservation was machine-readable at the relevant time of use, the second half of 2021; the court held the use covered, in addition, by the research exception in § 60d(1) UrhG (Federal Court of Justice case notice, checked 2026-11-10). The Federal Court of Justice held its hearing on 2026-09-03.
Until the pronouncement on 2026-12-17, the appellate reading on record is the OLG Hamburg judgment, and the form question stays open at the highest German court.
Does Training an AI Model Fall Within the Exception?
No court has yet ruled finally on whether training an AI model is text and data mining under Article 4, but the Advocate General in C-250/25 and two Munich judgments point in the same direction for the training copies and apart for what the model retains. In Like Company v Google (request lodged 2025-04-03 by the Budapest Környéki Törvényszék, concerning Gemini use claimed for 2023-06-13 to 2024-05-31), Advocate General Szpunar proposed on 2026-11-03 that the training questions are inadmissible as hypothetical (press release 155/26). For completeness he added that training is as a rule permitted under the exception. The opinion is not binding and the judgment is pending.
In GEMA v OpenAI (Munich I Regional Court, 42 O 14139/24, 2025-11-11, nine song lyrics), the court held that § 44b UrhG covers copies made to assemble a training corpus but not reproductions memorized in the model. The judgment is not final, and no higher-court docket is recorded on a court source. The AI copyright lawsuits over training data are tracked in one place.
In GEMA v Suno (42 O 763/25, 2026-07-31, six works), the same court held that memorization in models stored on servers in Germany infringes § 16 UrhG and that § 44b does not apply, while US training copies are governed by US law. That judgment is not final either. Corpus copies and model contents are therefore separate questions.
Why GPAI Providers Must Honor Reservations Under the AI Act
Regulation (EU) 2024/1689 makes Article 4(3) reservations a compliance duty for providers of general-purpose AI models: Article 53(1)(c) requires a copyright policy that identifies and complies with them, including through state-of-the-art technologies. The duty applies from 2025-08-02, AI Office enforcement starts on 2026-08-02, and models placed on the market before 2025-08-02 must comply by 2027-08-02. The open-source exemption in Article 53(2) does not cover point (c).
The GPAI Code of Practice (Copyright chapter, 2025-07-10) is voluntary. Its Measure 1.3 commits signatories to crawlers that read and follow robots.txt under RFC 9309 and other appropriate machine-readable protocols. SpaceXAI's training-content summary for Grok 4.7 states that it respects instructions that crawled content should not be used for training (a provider's own claim, as of 2026-09-21). The summary and template are covered under EU AI Act Article 53 duties.
Excluding Article 4 by Contract and Licensing Reserved Content
Article 7(1) makes contract terms contrary to Articles 3, 5 and 6 unenforceable but does not list Article 4, so a license or terms of access can exclude or narrow commercial mining. Content reserved under Article 4(3), or delivered under a license that excludes Article 4, can therefore be used for training only on the license's terms. The reservation and the license together are what a buyer pays for.
Article 4(1) also reaches the database right (Directive 96/9/EC Arts. 5(a) and 7(1)), and the sui generis right may itself be licensed (Art. 7(3)). Terms for mining and training are set out in the AI data licensing agreements guide.
Article 12 lets member states provide extended collective licensing, and each rightholder can exclude works at any time (Art. 12(3)(c)). This page records no member state's use of Article 12 for AI training.
Reserve, License or Do Nothing: Benefits, Risks, Downsides and Alternatives
For a rightholder in the EU, a reservation is the switch that turns free mining into a licensing question, but it costs effort and does not settle every case, and leaving content unreserved or not licensing at all remain valid choices.
Benefits
- A reservation plus a license makes training on the content a paid use, because Article 4 no longer applies (Art. 4(3), Art. 7(1)).
- GPAI providers must honor reservations under Art. 53(1)(c) of the AI Act.
- A license can set terms the exception cannot, such as excluding Article 4, as covered in should you sell or license your data to AI companies.
Risks
- The form of a natural-language reservation is untested at the Federal Court of Justice until 2026-12-17.
- The Advocate General reads training as a rule within the exception, and the judgment is pending.
- A reservation acts only where EU law applies, and the Munich court held US training copies subject to US law.
Downsides
- Machine-readable reservations need upkeep across protocols, and some operators state that their user-initiated fetchers may not follow robots.txt.
- Article 3 research mining stays open whatever is reserved.
Alternatives
- Reserve and do not license.
- Leave content unreserved.
- License through a collective route where one exists.
- Do not sell or license at all.
What Is Still Open: Rulings, Review and Policy as of November 2026
Five dated items can change how Article 4 is read, and each row gives its next step and source as of November 10, 2026.
| Item | Court or body | Status | Next step | Source | Page |
|---|---|---|---|---|---|
| Like Company v Google, C-250/25 | CJEU | AG opinion 2026-11-03: training questions proposed inadmissible; obiter, training as a rule within the exception | judgment pending | curia.europa.eu (press release 155/26) | Like Company v Google, C-250/25 |
| Kneschke v. LAION, I ZR 281/25 | Federal Court of Justice | hearing 2026-09-03 | pronouncement 2026-12-17 08:45 | bundesgerichtshof.de | Kneschke v. LAION, I ZR 281/25 |
| GEMA v. OpenAI, 42 O 14139/24 | Munich I Regional Court | judgment 2025-11-11, not final | no higher-court docket recorded on a court source | justiz.bayern.de | GEMA v. OpenAI, 42 O 14139/24 |
| GEMA v. Suno, 42 O 763/25 | Munich I Regional Court | judgment 2026-07-31, not final | appeal status not checked | justiz.bayern.de | GEMA v. Suno, 42 O 763/25 |
| Review of Directive (EU) 2019/790 | European Commission | review no sooner than 2026-06-07 (Art. 30(1)) | Commission review | eur-lex.europa.eu | n/a |
The European Parliament resolution P10_TA(2026)0066 of 2026-03-10 is a policy signal, not law: paragraph 10 asks for tools to exclude works from training, with the EUIPO as a trusted intermediary. No review outcome under Article 30 is recorded as of 2026-11-10.
Beyond Article 4: Other Jurisdictions, National Wording and the Seller Process
Article 4 is EU law only: the UK allows mining without a license only for non-commercial research under CDPA 1988 s.29A, and its March 2026 report calls an opt-out exception "no longer the government's preferred way forward", while US law decides training copies under fair use (17 U.S.C. 107) case by case. National wording for all 27 member states and the research rules sit under EU law for AI training data, and the pages below carry the rest.
- Jurisdictions: AI training data law by jurisdiction, TDM exceptions and opt-outs compared across jurisdictions, the UK research-only exception, the US approach.
- National texts: § 44b and § 60d UrhG, the French opt-out and the pending bill, Italy's AI-specific mining article.
- Reservations and routes: the protocols in technical detail, the sui generis right in a dataset, collective routes to a license.
- Seller process: a rights audit checks that the holder owns what it reserves or licenses, the steps to sell or license company data to AI companies follow it, and options other than a sale are listed apart.
What This Page Does Not Cover
This page covers the EU exception as it applies to data acquired for AI training, not publishers' and libraries' text-mining access services for researchers or their APIs. Switching off AI training in a platform's account settings is a different question (see account settings, a different question).
How This Page Is Sourced and Dated
Every rule cites the Official Journal text, the national statute or the court's own record, with the date it was checked. The method is set out in how we source and date every rule, and this page is general information about EU copyright law, not legal advice for any specific deal.
Frequently Asked Questions
Is text mining legal in the EU?
Yes, for anyone with lawful access unless the rightholder reserved it (Art. 4), and for research bodies without an opt-out (Art. 3). Both exceptions come from Directive (EU) 2019/790, and a contract can exclude only the first.
Does robots.txt count as an opt-out under Article 4(3)?
It can be one machine-readable means, but no EU court has ruled on it. Robots.txt is not binding by itself, and the GPAI Code of Practice makes following it a voluntary commitment.
Does the EU text and data mining exception apply in the UK?
No; the UK exception covers non-commercial research only (s.29A). The EU Directive does not apply in UK law.
Does Article 4 cover works published before June 2021?
The Directive applies to works protected by national law on or after June 7, 2021, which includes older works still in protection on that date (Art. 26(1)). Publication date alone does not decide it.