The Music Industry’s Next AI Copyright Fight Targets Claude

Sony Music Publishing, Warner Chappell Music, and affiliated publishers have sued Anthropic and two of its founders, alleging that copyrighted compositions were illegally acquired and copied while the company developed Claude. The complaint, filed August 28 in the Northern District of California as case 5:26-cv-09217, names CEO Dario Amodei and co-founder Benjamin Mann alongside Anthropic.

These are allegations, not findings. Anthropic disputes the publishers’ claims and says it intends to defend itself robustly in court. That distinction matters because the 48-page complaint tells only the plaintiffs’ side of a case that has just begun.

One lawsuit, four different copyright theories

The filing is broader than a simple argument about AI training. It brings four counts: alleged direct infringement through torrenting against all three defendants; alleged contributory infringement through torrenting against Amodei and Mann; and claims against Anthropic for direct infringement and for removing or altering copyright management information.

The publishers allege that torrent archives obtained by Anthropic contained books with their lyrics and sheet music, that licensed lyric sites were scraped, that physical books were destructively scanned, and that additional datasets were collected. They further allege that compositions were copied during training and reproduced in Claude outputs. Anthropic has not accepted those characterizations.

The complaint identifies tens of thousands of compositions in connection with some claims and cites familiar songs as examples. It asks for a jury trial, statutory damages, destruction of allegedly infringing copies, and an accounting of Claude’s training data. Those requested remedies should not be confused with damages awarded or facts established by a court.

Why acquisition and training must be separated

The case arrives after the Bartz authors’ litigation exposed a legal fault line that is easy to flatten in headlines: obtaining a work and using it for model training are separate acts. A court can view a transformative training use differently from acquiring source material through piracy.

Sony and Warner’s publishers are leaning heavily on that divide. Their complaint alleges that millions of books were downloaded through BitTorrent collections and that some included protected compositions. It also seeks to connect acquisition decisions to Amodei and Mann personally. Whether the new plaintiffs can prove ownership, copying, personal liability, and the scope of any infringement will be contested in this case.

For AI developers, the practical lesson is already clear without predicting a verdict: provenance is part of the product. A training pipeline needs records showing where material came from, what rights attach to it, who approved its use, and how it moved through preprocessing and model development.

Music rights make the dispute unusually layered

A song is not one indivisible asset. Lyrics, musical composition, and sound recording can carry different rights held by different parties. The present plaintiffs are music publishers focused on compositions, not a single consolidated owner suing over every element of a recording.

That structure helps explain why Anthropic now faces multiple music-publishing cases involving different catalogs and legal theories. Sony and Warner Chappell join earlier actions involving Universal Music Publishing Group, Concord, ABKCO, BMG, and Round Hill. Parallel cases do not prove the claims, but they show the scale of the licensing problem that frontier AI companies and rights holders have failed to settle commercially.

The output claim raises a separate product question

The publishers also allege that Claude can generate verbatim or near-verbatim lyrics and that safeguards can be circumvented by re-prompting. That is distinct from the question of whether training itself infringes copyright. It concerns what a deployed product returns to users and whether protected expression can be extracted.

Responsible model operators need layered controls here: dataset audits, memorization testing, filters for location-based requests to reproduce lyrics, repeated-prompt detection, and incident review when safeguards fail. None of those controls resolves the underlying legal questions, but each reduces avoidable exposure and makes the product’s behavior easier to examine.

The real fight is over a workable licensing market

The publishers say they already license compositions for some AI uses and argue that Claude should have been built within similar permission structures. AI companies, meanwhile, have strong incentives to resist a precedent that makes broad model development dependent on negotiating work by work across fragmented catalogs.

That tension is bigger than Anthropic. If courts draw a hard distinction between lawful analysis and unlawful acquisition, the market will reward clean datasets, collective licensing systems, and auditable content pipelines. If output-based claims gain traction, model providers will need much stronger evidence that their systems do not act as substitutes for licensed lyric services.

The Sony-Warner complaint is not a verdict on Claude. It is another sign that “we trained on the open web” is no longer enough of an answer. In the next phase of generative AI, the provenance of data and the behavior of outputs will be judged as separate, measurable parts of the product.

← Back to SunMarc App Labs