Two federal judges rule AI training on books can be fair use
Judge William Alsup ruled on 23 June 2025 that training a language model on books was transformative fair use, while sending Anthropic’s acquisition of more than seven million pirated books to trial. Judge Vince Chhabria issued a narrower fair-use ruling for Meta two days later.
Why it mattered They were the first American decisions to address squarely whether training on copyrighted books is fair use, and gave developers and rights holders a partial framework to argue from.
Authors had been suing the makers of large language models on a single theory: that copying books into a training set is copying, and that copying without a license infringes. Two orders from the Northern District of California, issued two days apart, gave the first substantive answers.
On 23 June 2025, Judge William Alsup granted Anthropic partial summary judgment in Bartz v. Anthropic. Training a model on books to produce new writing was, he held, “spectacularly” transformative and therefore fair use. Buying print books and scanning them into an internal digital library was fair use as well, since the company had lawfully bought every copy it converted.
The same order refused Anthropic the rest. The company had also downloaded more than seven million books from pirate libraries and kept them. Alsup held that acquiring a work by piracy is not excused by a later transformative use, and set that question for trial. The two halves of the ruling separated the act of training from the way the training data was obtained, and only the first was protected.
On 25 June, Judge Vince Chhabria reached a similar conclusion for Meta in Kadrey v. Meta, but on narrower ground. The authors had lost, he wrote, because they failed to show that Meta’s use harmed the market for their books, and he cautioned that the ruling should not be read as approving every use of copyrighted material in AI training. A better-argued case with evidence of market harm might come out the other way.
What the pair established was a boundary rather than a permission. Training itself could be fair use; the sourcing of the corpus was a separate exposure, and an expensive one. Anthropic agreed ten weeks later to pay about $1.5 billion to settle the piracy claims Alsup had left standing.