Whether training a large language model on copyrighted books is fair use has been one of the central open questions in AI copyright law. Bartz v. Anthropic, decided on summary judgment by Judge William Alsup in the Northern District of California, was the first US court decision to answer it directly, and the answer came in two very different halves.
Two Sources of Training Books
The claim, brought by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, centred on how Anthropic assembled the book corpus used to train its Claude models. Some books had been legally acquired. A separate, much larger set, more than seven million titles, had been downloaded from so-called shadow libraries such as LibGen, sites well known for distributing pirated e-book copies without authorization.
Training on Legally Acquired Books Was Fair Use
On the books Anthropic had acquired legally, Judge Alsup held that using them to train an AI model was "exceedingly transformative," and therefore protected fair use. The reasoning treated the model training process itself, extracting patterns and capabilities from text rather than reproducing or distributing the works, as a fundamentally different use than the original expressive purpose of the books, the kind of transformation fair use doctrine is meant to protect.
Piracy Was a Separate Question, and Anthropic Lost It
The transformative-use finding did not extend to how Anthropic obtained the pirated portion of its training corpus. Judge Alsup denied Anthropic's motion for summary judgment on that question, and cast real doubt on whether downloading pirated copies could ever be justified as reasonably necessary to a transformative training use, however transformative the ultimate training itself might be.
The Result: A $1.5 Billion Settlement
Facing exposure on the piracy claim, the parties reached a proposed class-wide settlement, announced in August 2025, covering authors and publishers whose pirated books were part of the training corpus, reported at $1.5 billion, the largest reported copyright class-action settlement in US history.
The Takeaway for Anyone Building or Licensing AI Training Data
The decision draws a sharp practical line: transformative use of copyrighted material for AI training has real legal footing, but that footing depends entirely on how the training material was obtained. An otherwise strong fair use argument does not rescue a training pipeline built on pirated or unauthorized source copies, a distinction that matters directly for any Canadian business licensing content for AI training, or defending against a claim that its own works were used without permission.
We advise clients on both sides of AI training data questions, licensing content for AI use, and responding when a client's own copyrighted works may have been used without authorization.