U.S. Judges Back Anthropic, Meta on Fair-Use AI Book Training

Two federal judges in San Francisco issued back-to-back decisions this week that for the first time squarely apply U.S. copyright law to artificial-intelligence training, finding that the use of copyrighted books to build large-language models can qualify as fair use.

On 23 June, U.S. District Judge William Alsup ruled in Bartz v. Anthropic that feeding legally obtained books into Anthropic’s Claude model was “spectacularly transformative” and therefore protected by fair use. He nevertheless refused to dismiss claims that the start-up kept more than seven million pirated titles in an internal “library,” sending that portion of the case to trial and leaving Anthropic exposed to potential damages.

Two days later, Judge Vince Chhabria granted summary judgment to Meta Platforms in a lawsuit brought by 13 authors who said their works were copied to train the company’s Llama model. The judge said the plaintiffs “made the wrong arguments” and failed to show market harm, concluding that, on the record before him, Meta’s training was transformative fair use. He stressed that the decision is limited to those authors and does not give the industry carte blanche, noting unresolved allegations that Meta distributed pirated files via torrents.

Together, the opinions offer early but influential guidance for dozens of pending AI-copyright suits. They endorse the view that large-scale text ingestion can be lawful when it is demonstrably transformative, while warning that the manner in which training data are acquired and stored—and the economic impact on original works—remain key legal fault lines for generative-AI developers.