For two years the answer to whether AI companies could train on your catalogue was effectively yes, because nobody had told them otherwise in a form they were obliged to read. That changed this month.
The 60-second answer
In the EU, AI companies may train on publicly available work under a text and data mining exception, unless the rightsholder has reserved their rights in a machine-readable way. If you reserve, your work falls outside the exception and training on it for the EU market requires a licence. Providers of general-purpose AI models are now obliged to have a policy for identifying and respecting those reservations. The obligation started on 2 August 2025 for new models, and the AI Office's full enforcement powers began on 2 August 2026.
- The mechanism is Article 4 of the EU's 2019 Copyright Directive: a broad mining exception, with an opt-out for rightsholders who reserve rights by machine-readable means.
- Model providers must have a copyright policy and respect those reservations under the AI Act. Music Business Worldwide has a plain-language breakdown for the music business.
- Models already on the market before August 2025 have until 2 August 2027 to comply.
- France's collecting society Sacem has published a general reservation of rights statement that members can use rather than drafting their own.
- 2 Aug 2025
- Obligations begin for models placed on the market after this date (EU AI Act)
- 2 Aug 2026
- AI Office full enforcement powers apply (EU AI Act)
- 2 Aug 2027
- Deadline for models already on the market before Aug 2025 (EU AI Act)
What machine-readable actually means
It is less technical than it sounds. The point is that a reservation must be findable by software rather than only by a human reading your website. In practice that means metadata attached to the files, robots directives on the site hosting them, and a clear reservation in your terms. A German court has held that a reservation written in natural language, for example inside terms of use, can qualify, which lowers the bar considerably for an artist without a legal team.
A reservation is a statement, not a lawsuit. Making one costs nothing. Not making one is treated as permission.
What this does not do
- It does not apply worldwide. This is EU law. A model trained elsewhere for a non-EU market is a different legal question, and one that is still being argued.
- It does not undo training that already happened. Reservations are forward-looking.
- It does not stop anyone imitating your style. Copyright protects the recording and the composition, not a sound or a vibe.
- It does not remove AI tracks that impersonate you. That is a platform takedown, handled separately and faster.
The practical minimum
If you do nothing else: add a reservation of rights line to the terms of the site your music lives on, ask your distributor in writing what reservation they apply on your behalf and get the answer in writing, and check whether your collecting society has published a statement you can adopt. Those three take an afternoon and put you on the correct side of a default that currently runs against you.
Frequently asked
Can I stop AI companies training on my music?
In the EU you can reserve your rights in a machine-readable way, which removes your work from the text and data mining exception. Training on it for that market then requires a licence. Outside the EU the position varies and is still being litigated.
Does my distributor opt out for me?
Some do and some do not, and terms change. Ask in writing which reservation they apply and to which territories, and keep the reply.
Is a note on my website enough?
A German court has accepted a natural-language reservation in terms of use as machine-readable. Pairing that with file metadata and site directives is stronger than relying on any one of them.
Does this protect me from someone imitating my style?
No. Copyright covers your recordings and compositions, not a genre, production style or timbre. Style imitation is not what a reservation addresses.