Claude Hit With Suit Over 7M Pirated Books, 'Tens of Thousands' of Songs

For AI engineers and model developers, the complaint is not just legal noise; it is a rare public inventory of alleged training-data contamination. It accuses Anthropic of torrenting at least 7 million books from pirate libraries and scraping lyrics from legal and archival sources, then reproducing copyrighted text in Claude outputs—a core training-data governance failure that will ripple through model evaluation, red-teaming, and dataset audits.

On August 28, 2026, Sony Music Publishing and Warner Chappell filed a multibillion-dollar copyright infringement action in a California court against Anthropic, its CEO Dario Amodei, and co-founder Benjamin Mann. The publishers manage rights for songwriters and composers, and they allege that Anthropic misused “tens of thousands” of copyrighted musical works to train its Claude family of AI models. The complaint describes the conduct as “one of the largest and most blatant ongoing thefts of intellectual property in history.” It is among the most consequential legal challenges to generative AI training to date, in part because it names individual executives and draws on evidence from a separate authors’ lawsuit that Anthropic settled for $1.5 billion.

“

On August 28, 2026, Sony Music Publishing and Warner Chappell filed a multibillion-dollar copyright infringement action in a California court against Anthropic, its CEO Dario Amodei, and co-founder Benjamin Mann.

The complaint identifies well-known compositions including Mariah Carey’s All I Want for Christmas Is You, Ain’t No Mountain High Enough, Survivor’s Eye of the Tiger, Bon Jovi’s Livin’ on a Prayer, Leonard Cohen’s Hallelujah, Mark Ronson’s Uptown Funk, and Katy Perry’s California Gurls. The plaintiffs assert that Anthropic and Benjamin Mann engaged in “torrenting, scraping and downloading” lyrics and sheet music to build the training corpus, and that Claude reproduced copyrighted lyrics in answers to user prompts. The claim that Mann personally downloaded at least seven million copies of books from pirate websites, including song lyrics and sheet music, is particularly significant. Those figures are said to come from a prior U.S. authors’ case against Anthropic, which the company settled for $1.5 billion. The publishers also allege Anthropic collected song data by scraping legal lyrics platforms such as Musixmatch and LyricFind and obtaining datasets from archive sites.

This case sits at the intersection of copyright, technology, and executive liability. Many earlier generative-AI suits have targeted companies, but naming Amodei and Mann as individual defendants may change the risk calculus for AI founders and executives. Plaintiffs will likely use the prior authors’ settlement as a factual foundation, arguing that the scale of alleged infringement is already established by discovery in that case. The $1.5 billion settlement figure signals that Anthropic has already paid a substantial price for similar claims, which may weaken its ability to argue that the alleged harms are speculative or de minimis. It may also provide a benchmark for damages in the music publishers’ action, though the works, rights, and remedies differ.

The complaint’s dual focus on training and output is important. Training-data claims generally involve one set of fair use and licensing questions; output reproduction claims introduce direct infringement analysis of the model’s responses. The publishers’ allegation that Claude can reproduce lyrics in response to prompts could trigger injunctive relief, output filtering requirements, or royalties. Music publishers are likely to seek not just retrospective damages but also forward-looking licensing and technical controls, potentially making this a bellwether for how courts address memorization and regurgitation in large language models.