A judge split AI training from the piracy that fed it

The writer's problem

Three authors, Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, sued Anthropic over how it built the book library behind Claude. The dispute reached a decision on 23 June 2025, when Judge William Alsup of the Northern District of California ruled on cross-motions for summary judgment, in an order that the Authors Guild has posted in full. The problem before him was narrow: which of Anthropic's specific acts of copying could be excused as fair use, and which could not.

What the documents show

The order itself states that Anthropic digitized purchased print books and separately downloaded more than seven million books from pirate sites to build a permanent library, then used subsets of both sources to train Claude. Judge Alsup held that training the model on the books, however they were obtained, was fair use, and that converting already-purchased print copies into digital form for internal storage was also fair use, reasoning that this merely replaced a physical copy with a more convenient one. But he denied Anthropic summary judgment on the pirated copies retained in its central library, and the order sets that question for a trial on damages. The Authors Guild's own analysis of the ruling agrees on what the order holds while disputing its reasoning, calling the piracy finding correct but arguing the training and format-conversion holdings misapply prior case law; that critique is the guild's own advocacy position, not an appellate finding, since no higher court had yet reviewed the order as this entry was prepared.

The editorial choice

An author or publisher reading this order has to separate two questions the ruling itself separates: whether training on lawfully acquired text is fair use, which this order answers yes, and whether acquiring the text unlawfully in the first place is excused by a later transformative use, which this order answers no. Treating the ruling as a blanket win or loss for either side collapses a distinction the judge drew deliberately.

What stays with the author

The order itself did not decide what damages pirated-book plaintiffs would recover; that question was instead resolved through a later class settlement rather than a trial. The ruling also does not extend to any other AI developer's training practices; it is specific to the record Anthropic built in this case. A reader still has to check whether a similar claim about another company's training data was acquired lawfully before assuming the same reasoning applies.

• Does a summary of this ruling distinguish training-copy fair use from pirated-copy liability, or blur them together?
• Has the underlying trial on piracy damages concluded, or is it still pending?
• Is a cited critique of the order a court's later holding, or an advocacy group's disagreement with the reasoning?

The Bartz order is a fair-use ruling about specific facts in one company's training pipeline, not a general legal clearance for training language models on any book by any means.