The dispute headed for a California federal courtroom this week, with Concord Music Group and other publishers accusing Anthropic of training Claude on copyrighted lyrics without permission. The next hearing is set for Oct. 21.¹

The publishers’ case has two tracks. They say Claude was trained on lyrics from their artists and can generate material resembling existing songs. More pointedly, a separate complaint alleges Anthropic obtained pirated copies of 20,517 musical works from torrent sites for training. In one example cited in a court filing, a request for “a song about the death of Buddy Holly” allegedly prompted Claude to produce the lyrics of Don McLean’s “American Pie” verbatim.¹

For publishers, the alleged copying is not merely an abstract training-data dispute. Licensed lyric websites pay to republish song texts; if users can obtain lyrics free from an AI chatbot, the plaintiffs argue, that revenue stream is undercut. They are seeking as much as $150,000 for each alleged copyright violation, plus up to $25,000 where copyright notices were allegedly removed—or alternatively the profits they say were lost.¹

Anthropic frames the case differently. It argues that using copyrighted works to train AI can qualify as fair use when the resulting use is transformative. It also says Claude includes “copyright-protective” measures and that a user who manages to elicit a copy is the actor who made it happen: “The user does.”¹

That argument arrives after a costly warning. Anthropic recently settled a book-publishers case for $1.5 billion; while a judge found AI training could be fair use, the company’s earlier use of pirated books remained a major vulnerability. Publishers are now pointing to a Sept. 30 Third Circuit ruling against ROSS Intelligence’s fair-use defense, though the California court is not bound by it. The central question is narrowing: even if AI training can be transformative, can that defense survive alleged reliance on pirated source material?¹