On August 2 the EU's requirement to label AI-generated content came into force. On the same day the bloc's AI Office gained power to enforce the training and copyright duties that model makers have carried since August 2025. Music Business Worldwide looks at what that means for the companies that build music AI, the companies that use it, and the rightsholders whose catalogues feed it.
Builders such as Suno and Udio now have to mark generated output in a machine-readable way, through watermarks or metadata, unless the AI only assists with routine edits. Systems already on sale have until December 2. Providers must also tell people when they are talking to an AI, which covers chatbots, virtual artists and fan-engagement bots. General-purpose model makers must follow EU copyright law, honour rightsholders' opt-outs, and publish a summary of their training data, wherever the training took place.
Deployers, meaning labels, streaming services and anyone else publishing AI output, must clearly label deepfakes, and a cloned voice of a real artist counts. When the clone sits inside an obviously artistic or satirical work the label can be light, closer to a credit than a stamp across the track. Venues and festivals that use face recognition or crowd analytics also have to tell the people being scanned. The training-data and copyright rules fall on model makers, not on the businesses that use their tools.
Rightsholders get their leverage from the 2019 Copyright Directive's text-and-data-mining opt-out. Once they reserve their works, those works cannot be used for training for the EU market without a licence. Sony and Warner have already written to AI firms withdrawing consent. Industry bodies including IFPI, CISAC and IMPALA say the training-data summary template is too weak to give them real transparency, and the Act does not ban training on copyrighted work outright. That leaves rightsholders to watch for breaches of their opt-outs and to sue when they find them.
Fines for breaking the labelling rules go up to 15 million euros or 3% of global turnover, whichever is higher, and national regulators enforce them. The copyright duties carry the same ceiling, enforced by the AI Office. Older general-purpose models have until 2027 to comply. Because the AI Office can now open cases on complaints, and the published training summaries give courts something to examine, MBW's view is that enforcement has become real. Its conclusion is that the EU has not banned AI music or its training, but has made disclosure and penalties enforceable for the first time.
The piece places the rules next to other recent moves: GEMA's first-instance win against Suno in Munich, which reached training done outside Europe, though that court relied on German collecting-society law rather than the Act, and labelling schemes proposed by labels or already run by Deezer, Qobuz, Tidal, Apple Music, Spotify and YouTube. It also notes that Deezer now receives more fully AI-generated uploads than any other kind. Deezer says fully AI-made tracks passed half of its new uploads in June, close to 90,000 a day. Meanwhile the majors' US cases against Suno and Udio continue, with no fair-use ruling expected before 2027.
Sources · 1