To opt out of AI training in the EU, a rightsholder must expressly reserve its text and data mining (TDM) rights under Article 4(3) of the Digital Single Market (DSM) Directive. For content published online, the reservation must be made “in an appropriate manner, such as machine-readable means”. A TDM opt-out written only in plain language may not be enough: in the German LAION case, the appeal court found that the photographer had not shown that a website’s natural-language ban on bots was machine-readable at the time. This guide is for publishers, photographers, brands and any business whose content is online.
Update (October 2026): the Federal Court of Justice (BGH) heard the LAION appeal on 3 September 2026. According to beck-aktuell, the court indicated that it sees a need for clarification under EU law and could refer questions to the Court of Justice of the EU; no date for a decision was announced. beck-aktuell (3 September 2026).
Key takeaways
- Article 4 of the DSM Directive allows TDM on lawfully accessible content unless the rightsholder has expressly reserved it; for online content, machine-readable means are the expected route.
- Spain transposed the rule in Article 67 of Royal Decree-law 24/2021.
- In the LAION case, the lower courts dismissed the photographer’s claim; the BGH scheduled its hearing for 3 September 2026.
- Since 2 August 2025, providers of general-purpose AI models must have a policy to identify and respect valid reservations (AI Act, Article 53(1)(c)).
- A robust opt-out uses several layers: robots.txt, metadata, terms of use and dated records.
What is the TDM exception and how does the opt-out work?
TDM means the automated analysis of digital content to extract information such as patterns, trends and correlations. The DSM Directive (EU) 2019/790 contains two exceptions. Article 3 covers research organisations and cultural heritage institutions for scientific research, and cannot be overridden by an opt-out. Article 4 covers everyone else, including commercial AI developers, but applies only “on condition that” the use has not been expressly reserved.
Recital 18 explains what an appropriate reservation looks like for online content: “machine-readable means, including metadata and terms and conditions of a website or a service”. For content not online, contracts or a unilateral declaration can work. In Spain, Article 67 of Royal Decree-law 24/2021 says the exception does not apply where rightsholders have expressly reserved use “to machine-reading means or other appropriate means”.
What did the German courts decide in the LAION case?
The facts, as summarised in the BGH press release of 20 May 2026 (case I ZR 281/25), are simple. LAION, a non-profit association, publishes a free dataset of 5.85 billion image-text pairs used to train generative AI. In 2021 it downloaded images to check that each matched its description; one was a preview of the claimant’s photograph from a stock agency’s website, whose terms prohibited access by “automated programs, applets, bots or the like”.
| Court | Date | Outcome |
|---|---|---|
| Hamburg Regional Court (310 O 227/23) | 27 September 2024 | Claim dismissed |
| Hamburg Higher Regional Court (5 U 104/24) | 10 December 2025 | Appeal dismissed: the copy was a reproduction, but covered by the TDM exceptions; the claimant had not shown the natural-language reservation was machine-readable in the second half of 2021; the research exception also applied |
| Federal Court of Justice (I ZR 281/25) | Hearing set for 3 September 2026 | Pending at the date of this article |
The lesson for rightsholders does not depend on the final outcome. The reservation was in the agency’s terms, written for humans, and the burden of proving it was machine-readable at the relevant time fell on the claimant. A rightsholder who also uses technical signals and keeps dated evidence is in a much stronger position.
How does the AI Act make TDM opt-outs matter?
Article 53(1)(c) of the AI Act requires providers of general-purpose AI models to put in place a policy to comply with EU copyright law and, in particular, to identify and comply with reservations made under Article 4(3), “including through state-of-the-art technologies”. Article 53(1)(d) adds a public summary of the content used for training. These obligations have applied since 2 August 2025.
The copyright chapter of the General-Purpose AI Code of Practice (10 July 2025) shows what signatories do in practice. They commit to:
- use web crawlers that read and follow robots.txt, as specified in IETF RFC 9309;
- identify and comply with other appropriate machine-readable protocols, such as asset-based or location-based metadata, that are standardised or widely adopted;
- publish information about their crawlers and how they respect reservations;
- offer a point of contact and a complaints mechanism for rightsholders.
Commitments of this kind turn a well-made reservation into something you can check and complain about. A reservation that crawlers cannot read gives you little to work with.
How to set up a TDM opt-out for AI training
| Layer | What to do | Why it matters |
|---|---|---|
| robots.txt | Disallow the crawlers used for AI training on the relevant paths, and review the list periodically | The protocol signatories of the GPAI code commit to follow |
| Metadata | Add a rights reservation to file metadata (images, documents, audio) and page-level metadata where supported | Travels with the file when it is copied elsewhere |
| Terms of use | State clearly that TDM under Article 4(3) is reserved, and say how to request a licence | Recognised by recital 18; useful but, after LAION, not enough on its own |
| Contracts | Require distributors, agencies and platforms that host your content to keep the reservation | Your content often sits on sites you do not control |
| Evidence | Keep dated copies of robots.txt, metadata and terms, for example through regular archiving | The burden of proving the reservation existed falls on you |
| Monitoring | Review crawler information and training summaries published by AI providers, and use their complaint channels | Turns the reservation into enforcement |
What this means for your business
- Decide your strategy first: blanket opt-out, opt-out with licensing, or allow certain uses. A reservation is also a bargaining tool.
- Implement the layers above on every domain and platform where your content appears, including country sites in Latin America and Africa.
- Assign responsibility: someone must maintain robots.txt and metadata as new crawlers appear.
- Keep records from day one; a reservation added later does not cover copies already made.
- Review licences that already allow third parties to use your content, so they do not undercut your opt-out.
Our AI training opt-out and digital assets service can design the reservation, draft the terms and licences and set up the evidence routine, alongside our copyright protection team.
Where companies get the TDM opt-out wrong
- Relying only on a sentence in the terms of use. The LAION appeal shows how hard it can be to prove that natural language was machine-readable.
- Blocking one crawler and forgetting others, or blocking search engines by mistake.
- Reserving rights on the main site but not on stock agencies, marketplaces or partner sites where the same content is published.
- Keeping no dated evidence of when the reservation was in place.
- Expecting the opt-out to cover research: the Article 3 exception for scientific research cannot be overridden. Where you believe a reservation was ignored, our IP enforcement team can assess the claim.
Frequently asked questions
Is a statement in my website’s terms of use enough to opt out of AI training?
Possibly, but it is risky on its own. Recital 18 of the DSM Directive mentions terms and conditions as machine-readable means, yet in the LAION case the Hamburg appeal court found that the claimant had not shown a natural-language reservation was machine-readable in 2021. Combine it with robots.txt, metadata and dated records.
Does the TDM opt-out stop all AI training on my content?
No. It stops reliance on the general exception in Article 4 of the DSM Directive. Research organisations and cultural heritage institutions can still rely on the Article 3 exception for scientific research. The opt-out also does not undo copies made before it was in place.
Do AI developers have to respect my opt-out?
Providers of general-purpose AI models placed on the EU market must have a policy to identify and comply with valid reservations under Article 53(1)(c) of the AI Act. Signatories of the GPAI Code of Practice also commit to follow robots.txt and offer a complaints mechanism.
Can IP Global Guard set up and enforce our opt-out?
Yes. We design the reservation strategy, draft terms and licences, set up the evidence routine and handle complaints and enforcement. In Spain and before EUIPO and WIPO we act directly where our professionals are entitled and otherwise through qualified representatives, coordinating local correspondents in Latin America and Africa.
Protect your content with IP Global Guard
A TDM opt-out only works if it is readable by machines, applied wherever your content lives and backed by evidence. IP Global Guard, the IP services line of META Channel Corporation Limited, handles that work with one strategy and one billing relationship across more than 25 jurisdictions; see our coverage in Europe, Latin America and Africa. The group’s AI Act practice covers the regulatory side.
Send us the list of domains and platforms where your content is published and tell us whether you want to block or license AI training. We will set up the opt-out and the evidence trail from a single point of contact. Talk to our content protection team.
This article is general information, not legal advice, and reflects the position on its publication date.
Sources
- Bundesgerichtshof, press release 85/2026, hearing in I ZR 281/25 (20 May 2026)
- Directive (EU) 2019/790 on copyright in the Digital Single Market (17 April 2019), Article 4 and recital 18
- BOE, Royal Decree-law 24/2021 of 2 November, Article 67
- AI Act, Article 53 (obligations for providers of general-purpose AI models)
- European Commission, General-Purpose AI Code of Practice, copyright chapter (10 July 2025)
- beck-aktuell, report on the BGH hearing (3 September 2026)
