Most publisher terms of use carry a sentence like this one: content on this site may not be used for text and data mining or for training AI models. A lawyer wrote it. It reads like a legal statement. In December 2025 a German appeal court held that a reservation in that form doesn't work, because the crawler it was aimed at was never able to read it.
What the Hamburg appeal court held
The case is Kneschke v LAION. The Hanseatisches Oberlandesgericht (Higher Regional Court) Hamburg gave judgment on 10 December 2025 in case 5 U 104/24 and dismissed the appeal.
The first confirmed that LAION qualified as a research organisation and could rely on section 60d of the German Copyright Act, which implements Article 3 of the DSM Directive, the scientific research exception for text and data mining.
The second reaches beyond this case. The court held that the rights reservation on the source website was ineffective, because it wasn't machine readable. General terms of use and human readable disclaimers don't satisfy Article 4(3) DSM. A reservation against commercial text and data mining has to be expressed in a machine readable format, and the court named the kinds it had in mind: robots.txt, the TDM Reservation Protocol, or metadata. The reporting on the appeal judgment sets out the reasoning.
The first instance had leaned the other way
The Landgericht (Regional Court) Hamburg decided the case at first instance on 27 September 2024. It dismissed the claim under the scientific research exception, so it never had to decide the Article 4 question. What it did do was remark, in passing, that it was inclined to treat a reservation written in natural language as machine readable, depending on the state of the technology. That remark was obiter. It bound nobody, but plenty of people read it as reassurance, and the first instance analysis circulated widely.
If your opt-out language was drafted on the strength of that commentary, re-read it.
What machine readable means in practice
The court didn't invent a standard. It pointed at mechanisms that already exist for telling an automated system what it may do with a file:
- robots.txt at the root of the domain, with directives aimed at the crawlers doing the mining
- The TDM Reservation Protocol, which expresses a reservation in a structured, parseable form
- Metadata attached to the file itself, so the reservation travels with the asset rather than sitting on a page the file gets copied away from
None of this replaces the clause in your terms of use, which still tells a human reader where you stand. What the ruling says is that the clause on its own isn't what Article 4(3) is looking for.
The case is not final
The appeal court expressly allowed a further appeal to the Bundesgerichtshof (Federal Court of Justice). This is the first appellate ruling in Germany on the point, not the last word on it, and anyone telling you the question is now settled across the EU is going further than the judgment does.
Waiting isn't a neutral choice, though. A reservation you put in place today was in place when a crawl happened. One you add after a dispute starts wasn't.
The AI Act puts the duty on the model provider, not on you
Separately, Article 53(1)(c) of the EU AI Act requires providers of general purpose AI models to put in place a policy to identify and comply with a rights reservation made under Article 4(3) DSM. That obligation has applied since 2 August 2025.
It's tempting to read that as protection, and it isn't, quite. It's a duty on the model provider to look for a reservation and respect it. It doesn't create the reservation for you, and it isn't evidence you hold. If your reservation isn't in a form the provider's policy can detect, the obligation has nothing to bite on. Our summary of the EU AI Act covers the provider side.
An opt-out is a signal, not a record of the work
A reservation is a permission signal addressed to whoever is crawling, and it works only if it's in the format the law expects. It says nothing about what you made or when you made it.
| Compare | Machine readable reservation | Dated record of the work |
|---|---|---|
| Question it answers | May this material be mined | What existed, in what form, on what date |
| Where it lives | robots.txt, a reservation protocol, file metadata | An independent, verifiable record you hold |
| If it fails | The mining may fall inside the exception | You argue authorship from memory and file dates |
Those two failures are independent. A rightsholder can lose the opt-out argument, as happened here, and still need to prove that a particular photograph or article existed in a particular form on a particular date. That second question comes up in infringement claims and licensing talks regardless of AI.
What to do about your own material
The last step is a different job from the first three.
- Move the reservation out of the terms of use and into a machine readable form, and keep the terms of use clause too
- Put the reservation at file level where you can, not only at page level, so it survives the file being copied
- Note when each reservation went live, since the date it was in place is what matters against any given crawl
- Separately, create a dated and independently verifiable record of the work itself, at the point it's finished
Swiss Trust Layer handles that last one. A qualified electronic seal and timestamp applied to a file proves that this exact file existed at that exact time, verifiable by anyone without depending on your own systems. It isn't an opt-out and it won't change how a crawler behaves. It's the record you still need on the day the opt-out question goes against you.
