Sony Music Publishing and Warner Chappell sue Anthropic over 'tens of thousands' of songs, naming Dario Amodei personally as the company prepares to list

The most important number in the lawsuit Sony Music Publishing and Warner Chappell filed against Anthropic last Friday is not the damages figure. It is seven million — the number of copies of books the publishers allege Anthropic downloaded from pirate websites. That figure does not come from the publishers' own investigation. It comes from a case Anthropic has already lost and already paid for.

The complaint, filed on August 28 in a California court and first reported by the trade publication Music Business Worldwide, accuses Anthropic of "torrenting, scraping and downloading" lyrics from copyrighted songs to train its Claude family of models, and then reproducing those lyrics in responses to users. It names the company, chief executive Dario Amodei and co-founder Benjamin Mann as defendants. The publishers describe what they allege as "one of the largest and most blatant ongoing thefts of intellectual property in history".

Anthropic disputes it. "We disagree with the publishers' claims and we intend to defend ourselves robustly in court," a spokesperson said.

Why the plaintiffs are publishers, not labels

A song carries two separate copyrights. The sound recording — the particular performance you hear — is usually owned or controlled by a record label. The underlying composition, meaning the melody and the words, belongs to the songwriter, and is administered by a music publisher. Sony Music Publishing and Warner Chappell are publishers. They are not suing over recordings of anything.

That distinction explains the shape of the case. The alleged harm is not that Claude listened to Bon Jovi. It is that the lyrics to "Livin' on a Prayer" — a text — were ingested as training data and can be reproduced as text. Compositions named in the complaint include Mariah Carey's "All I Want for Christmas Is You", "Ain't No Mountain High Enough", Survivor's "Eye of the Tiger", Leonard Cohen's "Hallelujah", Mark Ronson's "Uptown Funk" and Katy Perry's "California Gurls".

It also explains why the case is more dangerous to a language model developer than a recordings case would be. Lyrics are text. A model trained on text can be prompted to emit text. The publishers are alleging both the input and the output — that Anthropic acquired the works unlawfully, and that Claude has reproduced them.

The number that came from the last case

The seven-million-books figure is the hinge. It originates in the class action brought against Anthropic by US authors, which the company settled for $1.5 billion — the largest copyright settlement in American history, approved this year.

What made that case consequential was not the money but the split in Judge William Alsup's summary judgment ruling in June 2025. He found that training a large language model on books was "exceedingly transformative" and therefore fair use — a significant win for the industry. But he declined to extend that protection to works the company had obtained from pirate libraries. Training was lawful. Acquisition, on those facts, was not.

Defendants specifically elected to harvest unauthorized copies of Music Publishers' works so that they would not have to pay Music Publishers a licensing fee to use their works in AI training and development.

That ruling reframed the entire dispute. It moved the argument off the question AI developers had prepared for — is training transformative? — and onto one they were far less ready to answer: where did the file come from? Provenance, not purpose, became the exposed flank. The publishers' complaint is built directly into that gap, alleging that the same pirated corpus contained lyrics and sheet music alongside the books.

Two counts, not one

The complaint does not rest on infringement alone. It also alleges that Anthropic stripped identifying information from the works while processing them, denying rights holders what the publishers call "valuable attribution". That is a distinct statutory claim concerning copyright management information, and it carries its own penalty schedule.

The damages arithmetic follows from that structure. The publishers seek up to $150,000 for each infringed work — the ceiling for statutory damages where infringement is found to be wilful — plus $25,000 for each removal or alteration of identifying data. The complaint refers to "tens of thousands" of compositions, implying a floor of at least 20,000 works. Twenty thousand works at the statutory maximum is $3 billion before the second count is added at all.

Those are ceilings, not forecasts. Statutory maxima are rarely awarded in full, cases of this size settle far more often than they are tried, and the publishers must still establish both wilfulness and the count of works. But the ceiling is the number that sets the negotiating floor, and both sides know the last comparable case ended at $1.5 billion.

Where else the lyrics allegedly came from

Piracy is not the only acquisition route alleged. The complaint also says Anthropic scraped lyrics from legitimate licensed sites including Musixmatch and LyricFind, and drew on datasets held on archive services such as Common Crawl.

That widens the case in a way worth noticing. Scraping a lawful, publicly reachable website is a materially different act from downloading a torrent of pirated books, and it is far closer to the conduct that the Alsup ruling treated favourably. By pleading both, the publishers give themselves a claim that does not collapse if the piracy allegation fails — but they also invite the defence to argue that the strongest count and the weakest have been bundled together to make the whole look worse than its parts.

There is also a market the publishers can point to, which strengthens their position considerably. Music publishing is one of the few corners of the copyright economy with a mature, functioning licensing infrastructure. Publishers license lyrics to karaoke firms, to streaming services that display words alongside a track, to advertisers and to the very lyric sites named in the complaint. A rate exists. A negotiating process exists. When a plaintiff can show that a licence was available and was not sought, the fair-use analysis becomes harder for a defendant, because the fourth statutory factor asks about harm to the market for the work — and a market that demonstrably exists is easier to show harm to than a hypothetical one.

The wider docket

Anthropic is not facing this alone or for the first time. The company has already drawn claims from Universal Music Group and Concord Music Group. Earlier this month the AI music generator Suno lost a copyright case in Germany brought by the licensing agency GEMA, which alleged unlawful use of songs including Boney M's "Daddy Cool" to train its system — a decision in a different jurisdiction, against a different kind of model, but pointing the same way.

Ed Newton-Rex, a British composer who campaigns on artists' copyright, said he was "not surprised" that rights holders were moving against a company with a "history of downloading and training their AI models on huge libraries".

The timing

The filing lands while Anthropic prepares a stock market listing that could value the company at around $2 trillion. That is not incidental. A pending listing changes the calculus on both sides of a lawsuit: it raises the cost of unresolved contingent liability in a prospectus, and it tells any plaintiff exactly when their leverage peaks.

It also explains the decision to name Amodei and Mann personally rather than suing the corporate entity alone. Individual defendants raise the temperature, complicate insurance and indemnification, and make the matter harder to characterise to investors as routine commercial litigation. Whether the personal claims survive a motion to dismiss is a separate question from whether including them was effective.

The argument that has not been settled

There is a real defence here, and it is not merely procedural. Lyrics are short. Copyright protects expression, and the quantity of protected expression in a four-line chorus is genuinely contested territory. A model that has read a lyric and a model that will output it on request are not the same thing, and courts have not yet drawn a stable line between memorisation and reproduction. Anthropic will also argue, as it did successfully before, that the training use itself is transformative.

But the durable point is that the industry's central legal question has quietly changed. For three years the fight was over whether machine learning on copyrighted work is fair use. One court has now said that it can be — and then made the answer nearly irrelevant by asking where the training corpus was obtained. Every large model was built from data assembled in an era when nobody expected to be asked for a receipt. This complaint is another demand for one.

This report is based on reporting by The Guardian and Music Business Worldwide on the complaint filed August 28, 2026, on the Anthropic spokesperson's statement, and on the public record of Bartz v. Anthropic, including Judge William Alsup's June 2025 summary judgment ruling and the $1.5 billion settlement approved this year. Allegations in the complaint are untested and have not been proven. The damages figures cited are the maximum statutory amounts sought by the plaintiffs, not awards. The number of works at issue has not been fixed by the court.