Judge hears Anthropic's fair use defense on Claude lyrics

A federal judge heard cross-motions for summary judgment on Wednesday in the music publishers' copyright case against Anthropic over the song lyrics used to train the Claude chatbot, then took the question under submission without ruling.

U.S. District Judge Eumi K. Lee of the Northern District of California, sitting in San Jose, gave no sign of how she would decide after several hours of argument, Courthouse News Service reported. Bloomberg Law reported that Lee came to the hearing with eight binders of case materials and spent much of her questioning on the potential harm Anthropic's copying could cause the publishers.

The case is Concord Music Group, Inc., et al. v. Anthropic PBC, Case No. 5:24-cv-03811-EKL-SVK. Universal Music Group, Concord and ABKCO sued Anthropic in October 2023 in Nashville, and the case later moved to the Northern District of California. The complaint alleges Claude was trained on the lyrics of hundreds of songs without permission, including works by Beyoncé, the Rolling Stones and the Beach Boys. The publishers' opposition filing puts the number at 499 works in suit.

Anthropic's attorney, Sonal Mehta of WilmerHale, argued that training the model on copyrighted works is transformative because it leads to new content, which she said puts the practice within the fair use doctrine. She told the court that the publishers had not identified any specific market harm from Claude or other large language models, and cited Universal Music Group's own statement that AI is "overwhelmingly a net positive" for the music industry.

Matt Oppenheim of Oppenheim + Zebrak, for the publishers, argued that Anthropic uses the lyrics to generate competing content that harms human songwriters. According to Reuters, he said Claude received nearly 120,000 requests to generate "synthetic" lyrics over a six-month period, and warned that large language models "could overwhelm the market with AI-generated content very easily."

Lee pressed both sides on market harm, one of the four factors courts weigh in a fair use case, and questioned the publishers' assertion that Claude's ability to compose lyrics threatens the market for their lyrics. When Oppenheim argued that protecting the lyrics would hand the publishers a monopoly nobody else could work around, Courthouse News reported that he answered: "Copyright is a monopoly by nature." Mehta countered that "there isn't anything in the record that shows doom for the music industry," and that the synthetic songs Claude generates are new material in their own right.

Both sides asked for summary judgment, meaning a ruling on the written record with no trial. The publishers want Anthropic to obtain a license if Claude is going to use the works.

Lee has ruled on the case before. In January 2025 she ordered Anthropic to maintain its existing copyright guardrails for Claude's input and output and apply them to future models, but she did not broadly stop training on lyrics. In March 2025 she denied the publishers a preliminary injunction, writing that “Publishers are essentially asking the court to define the contours of a licensing market for AI training where the threshold question of fair use remains unsettled.” That same month she dismissed all but one claim and allowed the publishers to refile. In October 2025 she denied Anthropic's motion to dismiss, finding the publishers had plausibly alleged that Anthropic is paid when users request lyrics and again when Claude generates output that copies them.

Two other Northern California judges have split on the same question. In the authors' case against Anthropic, now-retired Judge William Alsup found that training gave the books a new purpose but faulted the company for keeping pirated books in a central library, and that case settled for $1.5 billion in 2025. Two days later, Judge Vince Chhabria ruled for Meta in a similar lawsuit but warned that AI training would not be fair use in many circumstances.

A ruling in the case could shape what AI video and music tools may train on, what their terms allow, and how hard their filters clamp down on what users type. If fair use holds for training, the data costs behind those tools stay where they are; if it fails, training data gets more expensive and licenses come first. Either way, the outcome would not settle the question nationally, because Alsup and Chhabria point in different directions.

The cross-motions remain under submission before Judge Lee, and no date has been set for her ruling.