Copyright

Kadrey v. Meta: A Fair-Use Win That Reads Like a Plaintiffs' Brief

Two days after Bartz, Judge Chhabria also found AI training to be fair use — but went out of his way to say the result reflected a failure of advocacy, not a vindication of the practice. His 'market dilution' theory is the doctrine to watch.

June 22, 2026
An open book with its pages turning
The court called AI training on books transformative, but flagged a market-harm theory the plaintiffs left unproven. Shutterstock
Educational content, not legal advice. This article explains general legal concepts. It does not create an attorney–client relationship. For your specific situation, consult a licensed attorney.

Decided two days after Bartz v. Anthropic, Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025), reached the same destination by a conspicuously different road. Judge Vince Chhabria granted Meta summary judgment on the fair-use question as to the thirteen author-plaintiffs before him — a group that includes Richard Kadrey and Sarah Silverman — but the opinion is best read not as an endorsement of training-as-fair-use but as a roadmap, addressed to future plaintiffs, for how this defense might be defeated. It is a rare summary-judgment order that tells the prevailing party how lucky it was.

At a glance

  • Case: Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal.)
  • Decided: June 25, 2025 (order on cross-motions for partial summary judgment), Judge Vince Chhabria
  • Holding: On this record, Meta’s use of the plaintiffs’ books to train its Llama models was fair use — but only because these plaintiffs failed to develop the market-harm theory the court considered most promising
  • Status: Partial, interlocutory ruling; other claims survive; the fair-use order is generally not yet appealable

The holding and its hedges

On the first fair-use factor under 17 U.S.C. § 107, the court agreed with the now-emerging consensus that training a large language model on books is “highly transformative.” But the analysis turned on the fourth factor — the effect on the market for the works — and it is there that Judge Chhabria parted company with any triumphalist reading of his own result.

He granted judgment to Meta because these plaintiffs had not built the evidentiary record the fourth factor demands. They showed no evidence that Llama reproduced their works in its output, and they did not develop the theory of market harm the court found most plausible. The opinion states the limitation in unusually candid terms: the ruling does not establish that training on copyrighted books is lawful generally, “only that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.” A companion order rejected the plaintiffs’ claim under 17 U.S.C. § 1202 of the DMCA, which concerns the removal or alteration of copyright-management information.

Market dilution as the live frontier

The doctrinal contribution of Kadrey is its articulation of what Judge Chhabria called a “market dilution” theory. The idea is this: a model trained on an author’s works can be used to generate a flood of machine-authored competition, depressing demand for human-authored books and thereby inflicting cognizable harm on the market under the fourth factor. The court flagged this as potentially “far more promising” than the theories actually advanced — and then, because it had been neither argued nor substantiated, declined to rest its decision on it.

That is a significant signal, and it is worth situating doctrinally. Since Harper & Row the fourth factor has been treated as among the most important, and Warhol reaffirmed the centrality of market substitution to the fair-use calculus. By naming market dilution as the theory most likely to carry the day, and by faulting the plaintiffs for neglecting it, the court has in effect published the prevailing theory of the next case. The conventional fourth-factor inquiry asks whether the secondary use substitutes for the original work; the dilution theory asks a subtler question — whether the secondary use floods the market the original competes in, even without copying the work itself. Whether courts will accept that broader conception of market harm is the open question on which the next round of AI-copyright litigation may turn.

Reading Bartz and Kadrey together

Taken in tandem, the two Northern District of California decisions establish that, on the records presented, training can qualify as transformative fair use — but neither should be mistaken for a categorical safe harbor. Bartz locates the residual risk in data acquisition: lawful use does not cure unlawful sourcing. Kadrey locates it in market harm and in the quality of the plaintiffs’ proof. The two are complementary maps of where AI-copyright liability still lives.

Both rulings are interlocutory, fact-bound, and non-precedential beyond their own dockets. The fair-use determination in Kadrey is unlikely to be appealable until the case resolves, and the opinion expressly disclaims any general holding. Practitioners should resist the temptation to cite either case for more than it decides.

Open questions

First, will any court adopt the market-dilution theory, and if so, what evidence will establish it — econometric proof of substitution, evidence of a displaced licensing market, or something else? Second, how will courts treat output that reproduces or closely paraphrases training works, a fact pattern Kadrey’s plaintiffs did not establish but a future plaintiff might? Third, the surviving claims in the case — including allegations tied to Meta’s acquisition of works from pirated sources — echo the acquisition theory that drove the result in Bartz, and their resolution may matter more than the fair-use headline.

Implications for businesses and creators

  • For rightsholders: The path to defeating an AI fair-use defense runs through the fourth factor. Build the record early — evidence of market substitution, of a licensing market the developer bypassed, and of any verbatim output.
  • For developers: A favorable training-fair-use ruling on one record does not generalize. The combination of Bartz (provenance) and Kadrey (market harm) defines the two fronts on which liability remains live.
  • For both: These are district-court rulings. The appellate law of AI and fair use has not yet been written.

Frequently asked questions

Did this case dismiss the authors’ lawsuit? No. It was a partial summary judgment limited to the training/fair-use and DMCA issues; other claims survive. Describing it as a dismissal of the lawsuit is imprecise.

What is the “market dilution” theory? The argument that AI output trained on an author’s work can saturate the market and depress demand for human-authored works, causing market harm under the fourth fair-use factor even without direct copying. The court called it promising but did not decide on it.

Does Kadrey mean training on copyrighted books is legal? No. The court expressly said it held only that these plaintiffs made the wrong arguments and failed to build the right record.

Authorities and sources