Meta won on fair use because authors could not show harm

The writer's problem

Thirteen authors, including Richard Kadrey and Sarah Silverman, sued Meta in a case whose own docket shows was filed 7 July 2023, over books used to train its Llama models. On 25 June 2025, Judge Vince Chhabria of the Northern District of California ruled on cross-motions for summary judgment, in an order denying the authors' motion and granting Meta's. The problem the order addresses is not whether copying occurred, which it treats as established, but whether the authors proved the kind of harm fair-use doctrine requires.

What the documents show

The order states Meta's use was highly transformative, which under the doctrine's fourth factor meant the authors needed to win decisively on market harm to defeat a fair-use defense. The judge distinguished three theories of harm: lost licensing fees for training data, which the order says is not a cognizable harm because authors are not legally entitled to monopolize a licensing market that would only exist to license the disputed use itself; reproduction of text by Llama's outputs, which the plaintiffs did not make their central claim; and market dilution, meaning AI-generated books flooding the market and competing with human-written ones. The order calls dilution the most serious potential harm and says it could often decide such cases for plaintiffs, but holds that these thirteen plaintiffs never raised the theory in their complaint or presented supporting evidence, leaving Meta's uncontested expert testimony that Llama's release had no measurable effect on their book sales unrebutted. That gap in the record, not a finding that AI training is generally harmless, is what the order says decided the case.

The editorial choice

Coverage that reads this ruling as a general finding that training on books is fair use overstates what the order holds; the judge wrote that better-developed evidence of market dilution would likely favor plaintiffs in a future case. A contemporaneous Authors Guild account of the parallel Bartz v. Anthropic ruling draws the same distinction from the plaintiffs' side, noting both cases originated in pirate-site downloads even where the market-harm outcomes diverged. Editors weighing how to describe either ruling should keep the court's reasoning separate from its outcome for the particular plaintiffs in front of it.

What stays with the author

An author's own evidence of declining sales, or the absence of it, remains central to any future claim of this kind; the order does not supply that evidence for anyone, and it does not extend its holding to plaintiffs who plead and support a market-dilution theory. The order leaves Meta's separately alleged distribution of pirated files, addressed elsewhere in the docket, unresolved here.

• Did the cited authors actually plead a market-dilution theory, or only raise it late in the case?
• Does a summary distinguish this ruling's evidentiary gap from a general finding about AI training?
• How does this order's market-harm reasoning compare with the piracy-specific holding in Bartz v. Anthropic?

The order is best read as a ruling about what one set of plaintiffs failed to prove, not as a verdict on whether AI training can ever harm the market for the books it trains on.