The AI Music Copyright Crisis

When I first wrote this essay, in November 2024, the honest description of AI and music copyright was a vacuum. The major labels had sued the two best-known song generators, no court anywhere had ruled on whether training a model on music is legal, and the European rules existed mostly on paper.

Two years later the vacuum has filled, partly with court rulings and partly with contracts. This is the state of play as of 29 October 2026, checked against the sources linked in the text. It is not legal advice. It is what I, as a working composer, think other composers need to know.

How I use and refuse these tools in my own studio is a separate question, and I wrote about it in The Impact of AI on Music Making.

1. Nov 2024Universal, Sony and Warner suing Suno and Udio in the US→Warner has settled with both and licensed its catalogue. Universal has settled with Udio. Universal and Sony are still suing Suno, Sony is still suing Udio, and both have filed second lawsuits.
2. Nov 2024No court ruling on AI training and music→Two judgments in Munich, against OpenAI over lyrics and against Suno over songs. Both say a licence is needed.
3. Nov 2024EU AI Act adopted, obligations still ahead→Copyright duties for model providers apply since August 2025 and can be fined since August 2026. AI-generated audio must carry a machine-readable mark.
4. Nov 2024Streaming services with no real AI policy→Deezer tags AI tracks and demonetises fraudulent streams. Spotify labels AI "artists" and carries AI credits. Bandcamp has banned AI-generated music.

The American lawsuits: from war to licences

In June 2024 Universal, Sony and Warner sued Suno in Massachusetts and Udio in New York, accusing both of training on their recordings without permission. A year later the front line had moved from the courtroom to the negotiating table.

• Universal and Udio settled at the end of October 2025 and signed licences for recordings and publishing. Udio promised a new subscription platform built on licensed music for 2026, and in the meantime stopped users downloading what they had generated.
• Warner and Udio settled on 19 November 2025, with a licensed service planned for 2026.
• Warner and Suno settled on 25 November 2025. Suno agreed to launch licensed models and retire the old ones, to limit downloads to paying users, and to let artists and songwriters opt in, or not, to the use of their names, likenesses, voices and compositions. Suno also bought the concert app Songkick from Warner.

The rest is still in court, and it has grown. In the Suno case the judge, F. Dennis Saylor IV, refused in August 2026 to add 61,026 more recordings to the original 560 because it would delay a case he wants decided on fair use. Universal and Sony then filed a second lawsuit on 18 September 2026 over 60,202 recordings. It targets Suno’s new v6 models, launched on 9 September and built, Suno says, with Warner, BMG and Believe. The labels call v6 “the fruit of the same poisoned tree”. Suno calls training on recordings “quintessential fair use”. In New York, Judge Alvin Hellerstein refused Sony’s request to add some 30,000 recordings to its Udio case on 29 June 2026, and Sony sued again on 20 July, pointing to its rivals’ licensing deals as proof that a market for training licences exists.

As of today, no American court has decided whether training a music generator is fair use. The nearest precedent is about books. In June 2025 a federal judge ruled that Anthropic’s training on lawfully acquired books was fair use, but that keeping millions of pirated copies was not. That case ended in a $1.5 billion settlement. This is why the labels insist on how the music was obtained: both new complaints allege the generators “stream-ripped” recordings from YouTube.

Germany: GEMA wins twice

The clearest answers so far have come from Munich, and they matter directly to anyone whose works are managed by GEMA, the German collecting society.

On 11 November 2025 the Munich Regional Court ruled that OpenAI had infringed the rights in the lyrics of nine German songs. The court found that the lyrics were memorised inside the models, that this memorisation is itself a copy, and that the text and data mining exception in EU law does not cover storing and reproducing whole works. OpenAI has appealed.

On 31 July 2026 the same court ruled largely for GEMA against Suno, over six compositions including “Forever Young”, “Atemlos”, “Big in Japan” and “Rasputin”. As analysed by Bristows, the court went further than in the OpenAI case:

• it found the works memorised in Suno models stored on German servers, and the outputs infringing, with Suno, not its users, responsible;
• it took jurisdiction over training that happened in the United States, ran its own US fair use analysis and rejected it;
• it refused the text and data mining exception, noting among other things that Suno had obtained audio by stream-ripping YouTube;
• it said that meeting the EU AI Act’s transparency duties does not remove liability for infringement.

The court granted injunctions and ordered Suno to disclose its revenue and pay damages, still to be quantified. The judgment is not final.

Europe: what the AI Act does and does not do

The AI Act does not decide whether training infringes copyright. Courts do that. What it does is force transparency and respect for the rights holder’s choice.

• Since 2 August 2025, providers of general-purpose AI models must have a policy to respect copyright, including rights reserved by rights holders against text and data mining under the 2019 Copyright Directive, and must publish a summary of their training content using the Commission’s template. A voluntary Code of Practice, published in July 2025, spells out how.
• Since 2 August 2026, the Commission can enforce those duties, with fines of up to 3% of worldwide turnover or €15 million, whichever is higher. Models already on the market before August 2025 have until August 2027.
• Since 2 August 2026, the transparency rules apply: providers of systems that generate synthetic audio, images, video or text must mark outputs in a machine-readable way so they can be detected as AI-generated. The Digital Omnibus, in force since 27 July 2026, gives systems already on the market until 2 December 2026 to add that marking. The Commission published a Code of Practice on marking and labelling on 10 June 2026.

The United States: authorship needs a human

The US Copyright Office has published two reports that matter to musicians. The second, in January 2025, concluded that prompts alone do not give the user enough control to be the author of the output. Human selection, arrangement and modification of AI material can still be protected; the generated material itself cannot. The third, on training, was released in May 2025 as a pre-publication version and says training can be fair use in some cases and not in others. The final version has not been published yet.

On 2 March 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the appeals court ruling that a work needs a human author to be protected.

The platforms: tagging, filtering, fraud

The streaming services are dealing less with copyright than with flood and fraud, but the two are now hard to separate.

• Deezer has tagged AI-generated tracks since June 2025 and keeps them out of algorithmic and editorial recommendations. In July 2026 it reported that fully AI-generated tracks had passed half of its new uploads at the June peak, close to 90,000 a day. They account for only 1 to 3% of streams, and Deezer found that up to 85% of the streams on them in 2025 were fraudulent and demonetised them.
• Spotify introduced in September 2025 a ban on unauthorised voice clones, a spam filter and AI disclosures in song credits. In its one-year review on 20 October 2026 it said it had taken more than 50 million slop and spam tracks out of recommendations this year, labels thousands of AI “personas” and keeps them out of recommendations by default, and that tracks likely to be AI-generated are ten times more likely to stay under its 1,000-stream royalty threshold.
• Bandcamp banned music generated wholly or substantially by AI in January 2026.

The fraud is real and now has a criminal precedent. On 6 October 2026 Michael Smith was sentenced in New York to 18 months in prison and ordered to pay back almost $8.1 million, after using bots to stream hundreds of thousands of AI-generated songs billions of times. I wrote about the fraud side in The Music Streaming Scam Epidemic.

What it means for a working composer

My reading of the last two years is simple. The idea that music on the internet is free raw material for anyone with a GPU has lost, in court and at the negotiating table. Training and output need licences, and licences mean money for the people who wrote the music. Whether that money reaches individual composers, rather than only the largest catalogues, depends partly on what we do.

In 2024 I ended this essay by saying the answer to all this cannot be creative paralysis. That still holds, with one difference. In 2024 it was a matter of faith. In 2026 the law, slowly and incompletely, is starting to agree that the people who make the music are the ones who should decide how it is used.