Anthropic, Meta Copyright Wins Hinge on Lawful Book Purchases

Two federal judges handed Anthropic and Meta partial victories in copyright cases, but the wins come attached to some very load-bearing footnotes. That is the argument Eryk Salvaggio makes in Tech Policy Press, and the framing is worth carrying into any celebration of the rulings.

Judge William Alsup, sitting in San Francisco, ruled that Anthropic did not violate copyright law by using books it had purchased to train its Claude models, likening the process to 'teaching school children how to write well' and calling it extremely transformative. He also noted that the authors 'concede that training LLMs did not result in any exact copies nor even infringing knockoffs of their works being provided to the public,' adding that 'If that were not so, this would be a different case.' The ruling only covered books Anthropic legally bought and digitized. A separate part of the same case, over allegations that Anthropic used pirated digital copies of millions of books, is due in December.

Judge Vince Chhabria's ruling in the Meta case landed in similar territory but with a sharper undertone. He treated training as transformative, but pushed back directly on the school-children analogy, writing that 'using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works.' Salvaggio's read is that Meta largely won on specificity grounds, while Chhabria quietly opened the door for tighter market-harm arguments the next time around.

Read together, these are narrow procedural wins that leave the door open on data provenance and on how future plaintiffs frame market harm. Salvaggio, who has argued that copyright will be an unsuccessful long-term lever against 'the AI industry's infringement of what are, fundamentally, data rights,' treats the rulings as roadmaps for future defeats rather than settled doctrine.

The honest caveat is that this is one outlet's analytical read, not the last word. The piece does not give you a granular breakdown of how much of Anthropic's dataset is actually pirated versus purchased, nor what a Meta appeal would look like, nor how the framework extends to non-book media like code and images. What is clear is where the pressure is moving. If you are running an AI training pipeline, the near-term implication is boringly practical and probably expensive. Provenance of your training corpus is now a legal artifact, and the next fight looks likely to be over outputs that resemble inputs, not over the act of ingesting them.