'Do not train' opt-outs for musicians: what works, what courts have said, and how to set one up

Our view, up front: a "do not train" opt-out is worth the half hour it takes, but it isn't a lock. For most musicians it works as a dated record of what you refused, and that record is useful mainly to whoever ends up enforcing it, which is usually your collecting society rather than you. Below is what the law allows, what courts have said so far, and a setup that is honest about where it stops working.

This is a practical guide, not legal advice. If you are weighing a claim, talk to a lawyer or your society.

What does EU law let you reserve?

The EU's 2019 copyright directive has two mining exceptions. Article 3 covers research organisations doing scientific research, and gives rightholders no opt-out. Article 4 covers everyone else, including commercial AI developers, but only if the rightholder hasn't "expressly reserved" the use "in an appropriate manner". For content online, the text names "machine-readable means".

Recital 18 is stricter. For content made publicly available online, it says a reservation should only be considered appropriate if it is machine-readable, "including metadata and terms and conditions of a website or a service". For anything else, such as a contract or a unilateral declaration, other means can work.

The European Commission's copyright-reform Q&A page refused our checker (HTTP 403), so this section relies on the directive text alone. For how the AI Act now ties model providers to these reservations, see our guide to the AI Act's training rules.

What have courts said so far?

Plain-language terms may not be enough. In Kneschke v LAION, a stock photographer's image was collected for LAION's dataset in 2021. The only reservation was written in plain language, in the website's terms of use and its source code, heise reports. On 10 December 2025 the Hamburg Higher Regional Court (5 U 104/24) dismissed his appeal. According to DLA Piper's summary, the court said a reservation must be interpretable by machines, and that "mere machine detectability is not sufficient". The photographer hadn't shown his reservation was machine-readable in the second half of 2021. That reverses the lower court's passing view that plain language would generally do. The court also found that no reservation is possible against the research exception, according to heise. A further appeal to the Federal Court of Justice was allowed, so this isn't settled.

Some copying is outside the exception altogether. In GEMA's case against OpenAI, the Munich Regional Court said on 11 November 2025 that the mining exceptions generally cover the preparatory copies made to build training data. But lyrics memorised inside the model were not text and data mining at all. The same chamber applied that reasoning to audio against Suno on 31 July 2026, as we covered in our report on the Munich ruling. Both rulings are first-instance and not final.

Put the two together and you can see where an opt-out matters. It doesn't decide whether memorisation is infringement, because Munich says that's infringing either way. It decides the step before: whether a commercial developer may copy your work to train on at all.

What about Spawning's Do Not Train Registry?

Spawning built the best-known private registry. In a 2024 guide it said the registry accepted any media with its own URL, music included, and that Stability AI and Hugging Face respected it. Those are Spawning's own claims, and we haven't confirmed them with either company. Shutterstock added over 444 million media URLs, the guide says. Spawning was also candid that opt-outs "are treated as voluntary in most jurisdictions".

There's a problem. When we checked on 23 October 2026, spawning.ai and HaveIBeenTrained both showed only an "Under Maintenance" page, and Spawning's blog has had no new posts since December 2024. We can't tell you whether the registry is still being read. That's the structural weakness of any private registry: your reservation is only as good as the company keeping it and the developers who choose to check it.

How to set up an opt-out that holds up

1. Block named training crawlers in robots.txt. Open your site's root folder (on hosted builders, look for an "SEO" or "robots.txt" setting) and add one block per crawler:

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: Google-Extended
Disallow: /

Each comes from the operator's own docs: OpenAI, Anthropic, Common Crawl and Google. Google says Google-Extended doesn't affect your inclusion in Search. OpenAI keeps its training and search crawlers separate. Blocking OAI-SearchBot keeps you out of ChatGPT search answers, so only block it if you mean to. OpenAI also says robots.txt rules may not apply to ChatGPT-User, because a person triggers those fetches. And the standard itself says these rules "are not a form of access authorization".

2. Add a written reservation as well. Put a short line in your site terms and in the description of each release you control, for example: "The rightholder reserves all rights in these works for text and data mining, including AI training, under Article 4(3) of Directive (EU) 2019/790." After Hamburg, treat this as a backup to step 1, not a replacement.

3. Consider ai.txt, knowing its limits. Spawning designed ai.txt to be checked when media is downloaded, not just when a page is crawled, with separate settings for audio and images. By Spawning's own account, it only covers files on your own site, and it depends on developers choosing to read it.

4. Ask your collecting society what it has reserved. Sacem announced an opt-out for its members' repertoire on 12 October 2023, requiring AI developers to seek prior authorisation. Sacem Luxembourg still carries the statement. The same commentary notes there is no standard way to opt out, and questions whether a society can opt out on members' behalf without their specific say-so.

5. Ask your distributor and label the same question in writing. Your audio mostly lives on streaming and video services. Spawning notes that registering your own domain doesn't cover your music embedded from SoundCloud or YouTube.

6. Keep dated proof. Screenshot your robots.txt and terms with the date, and keep the replies you get.

A two-line template for steps 4 and 5:

Please confirm in writing whether you have made a reservation under Article 4(3) of Directive (EU) 2019/790 for my works and recordings, where it is published, and in what machine-readable form. If you haven't, please tell me how I can make one.

Checklist

• robots.txt blocks GPTBot, ClaudeBot, CCBot and Google-Extended
• Written reservation in site terms and release notes
• Email sent to society, distributor and label
• Dated screenshots saved
• Calendar reminder to re-check crawler names every six months

What we think

Opt-outs fail musicians in a particular way: they only cover what you host, and only from now on. Anthropic, for example, says a ClaudeBot block signals that a site's future materials should be excluded from training. Recordings ripped from YouTube never pass through your robots.txt at all, and the Munich court's press release says that is how Suno got the songs in GEMA's case. Metadata doesn't save you either. Spawning found that X and Pinterest strip C2PA data when files are posted, a gap we look at in our guide to audio watermarks.

So the winners under this system are rightholders with scale. That means labels and societies that can reserve whole catalogues and then sue, as GEMA did. Independents mostly benefit through them. The UK's retreat from an opt-out exception suggests governments are hearing that opt-outs put the burden on the wrong side.

Do the six steps anyway. They cost little, and Hamburg shows the record matters. But put most of your energy into step 4. The question to ask your society isn't only "have you opted out?" It's also "who are you licensing to, and how will that money reach me?"