Sony Music, Warner Music, and other publishers have filed a federal lawsuit in Northern California accusing Anthropic of illegally torrenting tens of thousands of copyrighted song lyrics to train Claude. The 48-page complaint names CEO Dario Amodei and co-founder Benjamin Mann as individual defendants, alleging Amodei “expressly directed, approved, controlled, and intentionally induced these infringements.” The publishers are seeking up to $150,000 per infringed work, plus additional damages for stripping copyright management information.
The lawsuit deliberately targets the same weakness that cost Anthropic $1.5 billion in September 2025 — the largest copyright settlement in U.S. history. That case, involving pirated books, wasn’t lost because Anthropic used copyrighted material for training. It was lost because the company acquired that material through illegal torrent downloads from pirate libraries like LibGen and PiLiMi. This new complaint alleges Anthropic torrented at least seven million books from those sources, and treats the downloading itself as standalone infringement regardless of whether the content ever appeared in a commercial model.
Anthropic’s public denials about using pirated data hinge on a narrow definition of “training.” The publishers allege the company trained commercial Claude models on synthetic data generated by a non-commercial model that had learned directly from LibGen and PiLiMi texts. That non-commercial model allegedly also provided reinforcement feedback to commercial systems. If true, this is a classic laundering pattern — wash stolen text through an intermediate model, then claim the final product never touched the source. Discovery will reveal how much of that pipeline Anthropic documented internally.
Adding pressure, the Munich Regional Court ruled in November 2025 that song lyrics stored within a model’s parameters count as reproductions, and that chatbot output of those lyrics constitutes unlawful public disclosure. The German court placed responsibility squarely on model operators, not users who prompt for copyrighted lyrics. That ruling gives publishers a second legal front in Europe — and a template for arguing that the mere existence of lyrics inside Claude’s weights is itself an infringement, separate from any question of how training data was acquired.
