Every major label in the world has now had to decide what it thinks AI music is worth in court, and most of them decided the answer was a settlement cheque. Warner Music Group settled with Udio on 19 November 2025 and with Suno on 25 November 2025, six days apart. Universal Music Group settled with Udio on 29 October 2025 and announced its own licensed AI music platform for 2026, built on authorised recordings. The money changed hands, the complaints were withdrawn, and the industry looked like it was moving on.
Sony didn’t move on. Sony is still litigating against both Suno and Udio, and that decision matters more than any settlement. A licensing deal tells you what a catalogue is worth to an AI company trying to avoid the question. Whether training a model on copyrighted recordings without permission was lawful in the first place is a separate question, and a cheque doesn’t answer it. Sony is the only major forcing that question into a courtroom, and until a judge rules, every producer who has used a generative audio platform is building on ground with no legal foundation under it.
Discovery has already made the exposure concrete. When the RIAA filed against Suno in June 2024, the complaint listed 560 copyrighted works as an illustrative sample. Audio fingerprinting during discovery then matched millions of copyrighted recordings in Suno’s training data. In May 2026, UMG and Sony asked the court to add 61,026 recordings they had identified, describing that as only a fraction of the matches their experts found. Suno’s defence is that training AI models on copyrighted music is transformative use under US copyright law. That is the argument. No federal judge has ruled on it yet.
That ruling is not close. In UMG v. Suno, before Chief Judge F. Dennis Saylor IV in the District of Massachusetts, fact discovery closes on 30 September 2026 and dispositive motions are not due until 9 April 2027. So any definitive US decision on music-training fair use lands in 2027 at the earliest, and the precedent it will lean on cuts both ways. In June 2025, a Northern California court found in Bartz v. Anthropic that training a large language model on lawfully acquired books was fair use, calling it “exceedingly transformative,” but withheld that protection where the training material was piracy-sourced. The labels’ amended complaint against Suno, filed 19 September 2025, added a claim under DMCA §1201, alleging Suno built its training set by stream-ripping from YouTube and circumventing its rolling cipher encryption. If that allegation holds, Suno’s fair-use argument hits exactly the wall Bartz put up: lawful acquisition was the thing that saved Anthropic, and piracy was the thing that didn’t.
In the UK, the direction is clearer. On 18 March 2026, the government dropped a proposed text-and-data-mining exception that would have let AI firms train on copyrighted works without permission, after a campaign backed by Sir Paul McCartney, Dua Lipa, Sir Elton John, Björn Ulvaeus, Max Richter and Thom Yorke. Rights under the Copyright, Designs and Patents Act 1988 stay intact. Lawyers warn that existing copyright and data law still leaves voices and faces exposed to AI cloning, and the performers’ union Equity has been coordinating legal action and data-access requests on behalf of members whose voices were allegedly used without consent.
Spotify allows AI-generated music if the uploader holds the rights and the track doesn’t impersonate an identifiable artist. In the year to September 2025 it removed more than 75 million spam and fraudulent tracks under its spam and impersonation rules. The output itself has no protection to fall back on: in March 2026 the US Supreme Court declined to hear Thaler v. Perlmutter, leaving standing the rule that copyright requires a human author, so a purely machine-generated track is not eligible for copyright protection from the moment it exists. The output is unprotectable, the legality of the training data is unresolved, and the one company with the resources to force an answer is still in court. Using these platforms right now is not a creative risk. It is a legal one.
