The European Union‘s rules requiring AI-generated content to be labeled took effect on Sunday (August 2).
The same day, the EU‘s AI Office gained the power to enforce a separate set of obligations, in force since August 2025, that govern how AI models are trained on copyrighted work and what they must disclose.
Together the EU AI Act‘s rules now reach three groups across the music business: the companies that build AI, the companies that deploy it, and the rightsholders whose work trains it.
What the rules require for providers of AI
A provider is the company that builds an AI system or model, which in music means generators such as Suno and Udio.
Under Article 50, providers of generative systems must mark their output so it is machine-readable and detectable as artificially generated, using watermarks or embedded metadata.
That duty does not apply where the AI only assists with standard editing or does not “substantially alter” what a person supplied, and providers whose systems were already on the market have until December 2 to meet it.
Providers carry a second duty under Article 50, informing people when they are dealing with an AI system rather than a human unless that is obvious, which reaches chatbots, virtual artists and fan-engagement bots.
The Commission published guidelines on how the transparency rules apply in July.
Providers and deployers can also sign a voluntary Code of Practice on Transparency of AI-generated Content, published in June 2026, to show they meet the marking and labeling duties.
The makers of general-purpose AI models carry two further obligations that have applied since August 2025.
They must put in place a policy to comply with EU copyright law and to identify and respect the rights reservations, or opt-outs, that rightsholders attach to their work under Article 4(3) of the bloc’s 2019 Copyright Directive.
And they must publish a “sufficiently detailed summary” of the content used to train the model, on a template set by the AI Office.
Those duties bind any model offered in the EU regardless of where it was trained, so a developer cannot sidestep them by scraping and training outside the bloc.
What the rules mean for deployers of AI
Article 50 places a different duty on deployers, meaning whoever puts an AI system to use to create or publish content, from a label to a streaming service.
They must clearly and visibly label deepfakes, which the Commission defines as AI images, audio or video that resemble real people or events and would falsely appear authentic.
On the EU‘s definition the duty covers audio cloning a real artist’s voice, and it extends to AI-written text on public-interest matters that no human editor has reviewed.
Where such a clone is part of what the AI Act calls an “evidently artistic, creative, satirical, fictional or analogous work,” the disclosure is lighter.
Under Article 50 it need only be flagged in a way that “does not hamper the display or enjoyment of the work,” which points a label toward a credit rather than a mark across the track.
A further deployer duty under Article 50 covers emotion recognition and biometric categorization, reaching face-recognition entry and crowd analytics at EU venues and festivals, where those exposed must be told.
The Article 53 copyright and training-data rules do not fall on deployers; they sit with the model-makers.
What the rules mean for rightsholders
Rightsholders, meaning labels, publishers, songwriters and performers, draw their leverage from the opt-out written into the EU‘s 2019 Copyright Directive.
Under Article 4(3) of that directive they can reserve their works from text-and-data mining by machine-readable means, and where they do, a model cannot lawfully be trained on those tracks for the EU market without a license.
Sony Music and Warner Music Group have already written to AI companies withdrawing consent for their recordings and lyrics to be used in training.
The public training-data summaries are meant to give rightsholders a way to see what went into a model and, in the words of the AI Act, to “exercise and enforce their rights.”
A coalition including IFPI, CISAC, ICMP and IMPALA says that promise is going unmet.
Last summer, the group called the Commission‘s implementation of the copyright rules a “betrayal” of the Act’s objectives, saying its training-data template would not deliver “sufficient” transparency.
The EU rules force disclosure but do not ban training on copyrighted work, leaving rightsholders to police the opt-out and pursue infringement themselves.
The RIAA and IFPI have proposed a system for labeling AI tracks on streaming platforms as AI-generated or AI-assisted.
The major music companies, prominent indies and the IFPI have also proposed rules that bar such tracks from official charts unless they are lawfully made, substantially human-made, and flagged to listeners.
Universal Music Group and Sony Music remain in litigation with Suno in the United States, months after Warner Music Group settled with the company in late 2025 and entered a licensing deal.
In Europe, the Munich Regional Court ruled on July 31 that Suno had infringed copyright by training on six well-known compositions from the repertoire of German collecting society GEMA and reproducing them in its model and its outputs.
The court granted GEMA’s claims for an injunction, disclosure and damages: Suno must stop reproducing the works – including using them to train its models – and disclose the revenue earned so damages can be set. The ruling is first-instance and not yet final; Suno has said it is considering an appeal.
The court also reached the model training carried out in the United States. In a blog post published after the decision, law firm Reed Smith noted that the decision was “the first major European ruling” to find that unauthorized AI training “can infringe copyright even where the training itself takes place outside the EU”.
GEMA, in its statement on the win, said the decision means AI providers must license its repertoire even for training conducted outside Europe.
The cross-border reach is notable: it echoes the extraterritorial pull of the AI Act’s own copyright duties, which bite training wherever it happens – though the court grounded its jurisdiction in German collecting-society law, not the Act.
After the RIAA and IFPI proposed a system for labeling AI tracks last month, Suno, which the labels accuse of training on their music without permission, said “transparency is important” and that it is building watermarking and fingerprinting tools.
What happens if you don’t comply
Breaching the content-labeling rules can draw a fine of up to €15 million or 3% of a company’s worldwide annual turnover, whichever is higher, imposed by national regulators.
The copyright and training-data obligations carry the same ceiling, but there the Commission, through its AI Office, is the enforcer.
For a large model-maker or platform, the turnover share runs well beyond €15 million.
What platforms are doing already
The EU’s AI Act arrives as the volume of AI-generated music uploaded to streaming services continues to surge.
Paris-headquartered streaming service Deezer revealed last month that fully AI-generated music accounted for more than half of all new tracks uploaded to its platform for the first time in June, and it was receiving nearly 90,000 fully AI-generated tracks every day that same month.
Some streaming services have not waited for the law, including Deezer, which says it was the first streaming service to detect and tag AI music at the platform level, back in 2025.
High-fidelity rival Qobuz announced its own detection system in February, saying it would tag AI-generated tracks and remove those found to be impersonating artists or manipulating streaming activity.
YouTube also now labels photorealistic AI video even when a creator does not disclose it.
Spotify announced in September that it would support the new DDEX industry standard for AI disclosures in music credits. The company also began testing AI tags in its song credits in April, but only where an artist has chosen to disclose AI use through their label or distributor.
Spotify also introduced a new verification badge for artist profiles on its platform in April and said that “at launch, profiles that appear to primarily represent AI-generated or AI-persona artists are not eligible for verification”.
TIDAL set out a policy in June to tag tracks it identifies as fully AI-generated and stop them earning royalties.
Apple Music, meanwhile, launched a tagging system in March that relies on labels and distributors to declare AI-generated content, rather than detecting it at the platform level.
Amazon Music does not flag AI content.
What happens next
Providers whose general-purpose models were on the market before August 2025 have until 2027 to bring them into line with the copyright and training-data obligations.
Enforcement is no longer theoretical, with the AI Office now able to open cases on complaints and the public training-data summaries giving rightsholders and courts a document to point to.
And the labels’ lawsuits against Suno and Udio run on, with a US ruling on whether AI training is fair use not expected before 2027.
For all the detail, the EU has not banned AI music or its training; it has built the transparency and the penalties that make both enforceable for the first time.
