How to opt your music out of AI training: what EU law asks for, and what each signal actually controls

If you want your music kept out of commercial AI training in the EU, you have to say so, and you have to say it in a form a machine can read. A plain sentence in your website terms may not be enough; a named block in robots.txt plus a machine-readable reservation is the realistic minimum, and even that is a request, not a lock.

This is a practical guide, not legal advice. It covers EU law as we read it on 24 Oct 2026. Our August news piece, do-not-train opt-outs for musicians, covered registries and the court cases as news; this guide is the evergreen version: what each signal controls and in what order to set them up. For the licensing side, the deals AI companies sign when they do pay, see AI training data licensing explained.

What does EU law actually say?

The 2019 copyright directive, Directive (EU) 2019/790, has two text and data mining (TDM) exceptions. TDM is defined in Article 2(2) as "any automated analytical technique aimed at analysing text and data in digital form in order to generate information", which is broad enough that courts have been treating the copying done to build AI training sets under it.

• Article 3 covers scientific research by research organisations. It has no opt-out.
• Article 4 covers everyone else, including commercial developers. Article 4(3) says the exception applies only if the use "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". Article 4(4) says Article 4 does not affect Article 3.

Recital 18 adds the detail that matters. For content made publicly available online, "it should only be considered appropriate to reserve those rights by the use of machine-readable means, including metadata and terms and conditions of a website or a service." For content that is not online, "contractual agreements or a unilateral declaration" can work.

So the opt-out is real, but it only bites on the commercial exception, and the form of words is half the battle.

Does a sentence in my terms count?

Possibly not. In Kneschke v LAION, a photographer challenged a dataset built from a stock-photo site whose reservation was written in natural language. On 10 Dec 2025 the Hamburg Higher Regional Court (5 U 104/24) dismissed his appeal. According to DLA Piper's summary, the court held that a reservation must be capable of being interpreted by machines, that "mere machine detectability is not sufficient", and that the claimant had not shown the reservation was machine-readable in the second half of 2021. The first-instance court had taken the opposite view on natural language. The summary does not say whether robots.txt would have been enough.

Read that as: a written notice is useful evidence, but do not rely on it alone.

Why the AI Act raises the stakes

The EU AI Act ties the two together. Article 53(1)(c) requires providers of general-purpose AI models to "put in place a policy to comply with Union law on copyright" 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)". Article 53(1)(d) requires them to publish "a sufficiently detailed summary" of training content.

That turns your reservation from a private note into something a model provider is legally obliged to look for. It does not tell you which signal they will read, which is why we suggest sending more than one.

What each opt-out signal controls

Setting up robots.txt

Block training crawlers by name and leave search crawlers alone, so you stay findable. A minimal file:

User-agent: GPTBot
Disallow: /

User-agent: Google-Extended
Disallow: /

Both tokens come from the vendors' own pages: Google's crawler list and OpenAI's crawler overview. Other companies publish their own names; check each vendor's documentation rather than copying lists from forums, because names change.

Setting up TDMRep

The TDMRep report gives three ways to say "reserved": an HTTP header tdm-reservation: 1, which it calls the preferred technique; a JSON file at /.well-known/tdmrep.json; or an HTML tag, <meta name="tdm-reservation" content="1">. You can add tdm-policy, a URL to your licensing terms, so a developer knows how to ask. The spec advises using only one technique. On most artist sites, the meta tag is the one you can add without touching the server.

What none of this covers

• Platforms you upload to. Your songs mostly live on streaming services, video sites and stores. Their robots.txt is theirs, not yours, so your file protects your site and little else. Ask your distributor and your collecting society, in writing, what they have reserved on your behalf.
• Copying already done. A reservation speaks to future mining. Hamburg looked at the state of the site at the time of the 2021 crawl.
• Research use. Article 3 has no opt-out.
• What a model memorises. In GEMA v OpenAI (42 O 14139/24), the Munich Regional Court accepted on 11 Nov 2025 that TDM exceptions can cover preparatory copies made to assemble training data, but held that training which reproduces works in the model is not covered. The judgment is not final. That is a separate route to a claim, opt-out or not.

Our order of setup

1. Add named robots.txt blocks for the training crawlers you can verify.
2. Add one TDMRep signal, ideally with a tdm-policy link to a page that says how to license your catalogue.
3. Put a dated, plain-language reservation in your site terms and in your release notes.
4. Email your distributor and collecting society asking what reservation they make for your catalogue, and keep the replies.
5. Screenshot all of it with dates. If you ever need to show a reservation existed before a crawl, this is your record.

What we think

The opt-out is built backwards for musicians. The directive puts the burden on the rightholder to speak a machine's language, and Hamburg has now said a human sentence may not be enough. Most artists do not run the servers their music sits on, so the people best placed to reserve rights are the platforms, distributors and collecting societies, and they should be doing it by default and telling artists exactly what they reserved.

Until that happens, set the signals up anyway. They take an afternoon, they cost nothing, and under Article 53 they are now something model providers must look for. Treat them as a dated record and a licensing door, not as a wall. The real leverage is still the licence: a clear tdm-policy link that says "ask, and pay" is worth more than a block list. And if your concern is your voice rather than your recordings, read how to protect your voice from AI cloning.

FAQ

Does blocking Google-Extended hurt my search ranking? Google says it does not affect inclusion in Search and is not a ranking signal.

Is robots.txt legally binding? The standard itself says it is not access authorization. In the EU it may serve as a machine-readable reservation, but no court we found has confirmed that for robots.txt specifically.

Does an opt-out stop research labs? No. Article 3 research mining has no reservation.

Can I opt out of training that already happened? A reservation looks forward. It does not undo copies made before it existed.

Updated 24 Oct 2026.