Federal judge sharply questions $1.5 billion Anthropic settlement in pirated books case

A federal judge on Monday criticized a $1.5 billion settlement between artificial intelligence company Anthropic and authors who say nearly half a million books were illegally pirated to train AI chatbots, raising the possibility the accord could still be undone or the case sent to trial.

U.S. District Judge William Alsup spent nearly an hour at a San Francisco hearing expressing deep reservations about the proposed resolution and its claims process, and scheduled another court session for Sept. 25 to review whether his concerns have been addressed. He ordered a “drop-dead list” of the total books at issue by Sept. 15 and said the parties must submit a proposed claims form by Sept. 22 for his review.

The settlement, announced days before the hearing, would pay authors and publishers roughly $3,000 per book covered by the agreement and is intended to resolve claims that Anthropic trained its Claude chatbot using pirated works. Attorneys for the authors told the court that about 465,000 books are on the list of allegedly pirated works. Alsup said he needed stronger assurances that the number would not grow, warning that the company could be blindsided by additional claims.

Alsup, who issued a mixed ruling in June finding that using copyrighted books to train AI chatbots is not per se illegal while also concluding Anthropic had wrongfully acquired millions of books from pirate websites, said he was concerned that the settlement’s notice and claims procedures might leave eligible authors unaware and deprive them of a fair chance to participate. “We’ll see if I can hold my nose and approve it,” he said at the hearing.

The judge also questioned the roles of major industry groups involved in the talks, including the Authors Guild and the Association of American Publishers, saying he feared they might exert undue pressure on authors to accept the deal without fully understanding it. Both groups have been active in the case: Authors Guild CEO Mary Rasenberger and AAP CEO Maria Pallante attended Monday’s hearing.

Pallante called some of Alsup’s revised timetable for approving the settlement “troubling,” saying the judge “demonstrated a lack of understanding of how the publishing industry works.” In a statement after the hearing, the Authors Guild said it was “confused” by Alsup’s suggestion the organization might be secretly undermining some writers, adding that its work on the settlement aims to ensure authors’ interests are fully represented and is being conducted with transparency.

Plaintiffs in the underlying suit — authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson — were present in the courtroom but did not speak. Johnson had described the settlement previously as the “beginning of a fight on behalf of humans that don’t believe we have to sacrifice everything on the altar of AI.” Justin Nelson, an attorney for the authors, told the judge he and his colleagues were confident the funds would be fairly distributed, noting the case’s high-profile media coverage.

Alsup warned, however, that he was wary of “hangers on in the shadows” and said he might permit the case to proceed to trial if the settlement protections prove inadequate. The lawsuit had been set for trial in December before the settlement was announced.

Anthropic and the parties proposing the deal told the court they will respond to the judge’s demands and submit the required lists and claims materials by the deadlines. The Sept. 25 hearing will determine whether those revisions satisfy Alsup’s concerns and whether the settlement can move forward.

The case is among several recent lawsuits challenging how AI companies obtain and use copyrighted material to train models. Courts have been grappling with questions about data sources, fair use and the responsibilities of companies that develop generative AI systems, and rulings in these matters could shape industry practices and future litigation.

Reporting from Providence, Rhode Island, contributed to this story.