US Judges Hand Landmark Fair-Use Wins to Anthropic and Meta in AI Book Disputes

A pair of federal court decisions in San Francisco has given artificial-intelligence developers significant, if qualified, support for using copyrighted books to train large-language models. On 24 June, U.S. District Judge William Alsup ruled that Anthropic PBC’s practice of buying, scanning and feeding printed books into its Claude chatbot was “exceedingly transformative” and therefore protected by the fair-use doctrine. The opinion dismissed infringement claims tied to legally acquired works but left Anthropic facing trial over allegations that it also kept a digital library of some seven million pirated titles—exposure that plaintiffs say could run to more than $5 billion in damages.

Two days later, Judge Vince Chhabria threw out a separate lawsuit in which 13 authors accused Meta Platforms of using pirated versions of their books to train its Llama model. The judge said the writers “made the wrong arguments” and offered no evidence of market harm, but stressed that unlicensed copying could be unlawful in other circumstances and that his decision applied only to the plaintiffs before him.

The back-to-back rulings are the first to squarely apply fair-use principles to generative-AI training, offering an initial legal roadmap for dozens of pending copyright cases. Within hours of the decisions, a new lawsuit was filed in New York alleging Microsoft relied on around 200,000 pirated books to develop its Megatron system, underscoring the continuing uncertainty over how courts will balance creators’ rights against the data demands of AI research.