What it does This order holds that Meta’s copying was highly transformative, and that copying entire books was reasonably necessary to train an LLM, but it also recognizes that market dilution from AI-generated competing works could in other cases defeat fair use. The authors’ theories—that Llama could reproduce meaningful portions of their books and that unauthorized copying harmed a licensing market for AI training—failed on this record. Llama could produce no more than 50 words and punctuation marks from the books under adversarial prompting, and the plaintiffs offered no meaningful evidence of market dilution. The order does not decide Meta’s alleged distribution of works through torrenting; that issue remained live.

Who it affects The ruling resolves only the individual claims of the thirteen named authors, not the proposed class or other copyright holders whose works Meta used. It directly affects Meta, the Llama models, and authors whose books were among at least 666 downloaded copies.

Why it matters The decision is fact-specific rather than a blanket approval of AI training on copyrighted works. It states that LLM training may often fail fair use where better evidence shows substantial market harm, especially from indirect substitution or market flooding. If copying is not fair use, the court said developers may need licenses or public-domain works rather than abandon AI development.

Key dates and numbers

• Order dated June 25, 2025; case-management conference scheduled for July 11, 2025.
• Case No. 23-cv-03417-VC; order concerns Dkt. Nos. 482 and 501.
• Thirteen named authors; at least 666 copies of books in which they held copyrights were downloaded.
• Llama testing produced no more than 50 words and punctuation marks from the plaintiffs’ books.