Concord Music Group and other publishers sued Anthropic in 2023, alleging that Claude was trained on copyrighted lyrics and can reproduce protected songs or generate work too close to existing material. Their broader complaint is economic: lyric websites pay to display songs, while an AI answer can make that paid market look dispensable. In one cited test, publishers say Claude responded to a request for a song about Buddy Holly with a verbatim rendering of “American Pie.”¹

The dispute extends beyond model training. In a separate case, Concord alleges Anthropic obtained more than 20,000 music works from torrent sites. The publishers are seeking as much as $150,000 for each of 499 songs in the main lawsuit, plus damages tied to allegedly stripped copyright notices. They argue that free lyric outputs weaken their leverage to negotiate licenses with businesses that would otherwise have to pay.

At an Oct. 21 hearing in San Jose, California, Anthropic pressed a sharply different account. It says using lyrics within a vast training corpus is fair use, that Claude was not built as a copying machine, and that it has installed safeguards against identical reproductions. Attorney Sonal Mehta argued there was no evidence Claude had been specifically designed to output copyrighted lyrics, calling such a product “the world’s most expensive lyric-generating machine.”²

Anthropic further says a user who jailbreaks those protections—not the company—causes any unlawful copy. Publishers reply that key guardrails arrived only after litigation began. The hearing did not resolve the competing motions; Judge Eumi Lee asked the parties to consider additional dates. The eventual ruling will turn not simply on what Claude learned, but on whether it substitutes for the licensing market the publishers say copyright is meant to protect.