Presented by Kyle Coleman
Training an AI on books can be fair use — but downloading pirated copies to build a permanent library is not. That was the split ruling in Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal. June 23, 2025), the first court decision to address fair use for training a large language model. Judge William Alsup held that using the books to train Anthropic's Claude models was 'exceedingly transformative' and a fair use, and that converting purchased print books to digital form was also fair use — but that acquiring and retaining pirated copies for a central library was not excused, setting those copies for trial on liability and damages. The parties later settled before trial.
Bartz v. Anthropic, decided June 23rd, 2025, by Judge William Alsup of the Northern District of California — the first court decision to address fair use for training a large language model. The court held that using the books at issue to train Anthropic's Claude models was a fair use, and that converting purchased print books into digital files for a central library was also a fair use. But downloading and keeping pirated copies to build that library was not excused by fair use — and the court set those pirated copies for trial. Here's the brief.
Three authors — Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — sued Anthropic, alleging it had copied their books to build a central library and to train its Claude large language models.
Anthropic had obtained books two ways. It downloaded millions of copies from online pirate libraries. And it also purchased printed books, tore off the bindings, scanned the pages, and discarded the paper originals — creating digital copies in their place.
Anthropic moved for summary judgment, arguing that its copying was fair use under Section 107 of the Copyright Act. The authors did not contend that any text Claude generated for users reproduced their works; their claim was about the copying of the books themselves — to train the models and to stock the library.
The question was whether Anthropic's copying qualified as fair use.
Following the Supreme Court's decision in Warhol, the court did not treat the copying as a single act. It identified distinct uses and analyzed each on its own: the copies made to train the models; the digital copies made from purchased print books; and the pirated copies held in the central library.
On the training copies, the court found the first factor decisively in Anthropic's favor. Using the works to train the models, it reasoned, was transformative — "spectacularly so" — because the purpose was to build a tool that generates new text, not to supplant the books. Weighing all four factors, the court concluded the training use was fair.
The format change fared the same way, for a narrower reason: replacing a purchased print copy with a searchable digital copy — without adding copies, creating new works, or redistributing anything — was a fair use.
The pirated library copies were different. Building a permanent, general-purpose library, the court held, did not itself excuse acquiring the books by piracy.
Whether the pirated copies were later used for the transformative training, the court reasoned, did not retroactively cure the initial piracy.
Bartz v. Anthropic was the first decision to apply fair use to the training of a generative model, and it drew a line: the transformative training use was fair, but the acquisition of pirated copies to build a library was not, and would be tried, with damages including for willfulness. The ruling came at summary judgment on this record; the parties later settled before the scheduled trial.
Bartz v. Anthropic, decided June 23rd, 2025. I'm Kyle Coleman. Thanks for watching.
These videos are educational case briefs, not legal advice, and watching them does not create an attorney-client relationship with the presenter or the firm. Case law and its interpretation evolves, always check a decision's subsequent history. Do not rely on these case briefs, but read the case yourself or have your attorney read them. Videos are presented via an AI avatar and voice clone of Kyle Coleman, created with his participation and consent.