Two federal judges find training on books fair use on their records, and leave pirated libraries and market dilution open

2025-06-25Copyright

Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal. June 23, 2025) (Alsup, J.) (order on fair use); Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. June 25, 2025) (Chhabria, J.) (order denying the plaintiffs' motion for partial summary judgment and granting Meta's cross-motion for partial summary judgment)

Judge William Alsup granted Anthropic summary judgment that using books to train the large language models behind Claude was fair use under Section 107 of Title 17 of the United States Code, calling the training use "exceedingly transformative," and held that scanning purchased print books into digital library copies was fair use as well. He denied summary judgment for the more than seven million pirated copies Anthropic downloaded from Books3, Library Genesis and the Pirate Library Mirror for a permanent central library, holding that "Anthropic had no entitlement to use pirated copies for its central library," and set those copies for trial. The order noted that the authors alleged no infringing outputs and that the Claude service placed software between users and the model to block them. On June 25 Judge Vince Chhabria granted Meta summary judgment on fair use for training Llama on thirteen authors' books taken from shadow libraries, finding the use highly transformative, Llama "not capable of generating enough text from the plaintiffs' books to matter," and the authors not entitled to a market for licensing their works as training data. He called market dilution the potentially winning argument, found no evidence of it, and stated that the ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful."

What it changesBoth judges treated training as transformative. Judge Alsup then separated the training copies from the pirated library and sent the library to trial, and Judge Chhabria ruled on the absence of market evidence, so the records that decide these cases are how the copies were obtained and kept and what the model can reproduce. The acquisition record also feeds the patent program. A specification that names its training corpus is a public admission about where the data came from, so describe training data by its properties and processing and keep sources in the confidential provenance file. Keep that file per training run, with source, license, acquisition method and retention. The output filter that mattered in Bartz is engineering worth an invention disclosure, since blocking regurgitation is now both a technical feature and a litigation defense. These are district court rulings on particular records, and the Meta decision binds only its thirteen plaintiffs.