A judge rules training on books is fair use

Judge William Alsup of the Northern District of California ruled on summary judgment in Bartz v. Anthropic, a class action brought by authors, that training a large language model on lawfully acquired books was fair use, while denying Anthropic summary judgment on its use of pirated book copies — sending that narrower question toward trial. It was one of the first US court rulings to address squarely whether training an LLM on copyrighted text is fair use.

Working through the statute’s four fair-use factors, Alsup found training “exceedingly,” even “spectacularly,” transformative: the model was not reproducing the books but learning from them to generate new text, which he likened to training a student to write well rather than to plagiarise. He also found that Anthropic’s practice of buying print books, scanning them, and discarding the physical copies was fair use, treating the digital copies as replacements for books already lawfully owned rather than as additional infringing reproductions. On the fourth factor, market harm, he rejected the argument that training itself displaced demand for the authors’ works, comparing the claim to complaining that teaching children to write well would flood the market with competing writers.

That reasoning did not extend to books obtained through piracy and kept in what Anthropic called a permanent digitised library regardless of whether they were later used for training. Alsup found no purchase-and-replace justification applied to material Anthropic never bought, and held that downloading pirated copies did not become fair use merely because some of those copies were eventually used in a transformative training process. That question — how many pirated works were involved and what Anthropic owed for them — went toward trial rather than being resolved on summary judgment, and the case was ultimately resolved through a settlement covering hundreds of thousands of books. The split ruling gave AI developers a partial legal foundation for training on purchased or licensed text while leaving piracy-tainted datasets, a common shortcut in building training corpora, exposed to liability.