Kadrey v. Meta Platforms is a copyright action by thirteen authors, mostly fiction writers, who allege that Meta downloaded their books from online shadow libraries and used them to train its Llama large language models without permission or payment. On 2025-06-25 Judge Vince Chhabria granted Meta partial summary judgment that its use of these plaintiffs’ books as training data was fair use on the record before him, while stating that in most cases of this kind the answer would likely be the opposite.[1]
Background
The plaintiffs alleged that Llama and its outputs were infringing derivative works. In November 2023 the court dismissed those theories, calling the idea that the models are themselves derivative works “nonsensical” and finding no allegation of an infringing output.[3] The claim that copying the books for training was unauthorised copying continued.[3] How Meta licenses Llama itself is described in Meta Llama licensing.
The dispute
The summary judgment record described Meta’s attempts to license books. Meta’s head of generative AI discussed spending up to USD 100 million on licensing, but Meta found licensing harder than expected: publishers generally did not hold the rights to license books for AI training, there was no collective licensing organisation, and some publishers were unresponsive. After escalation within the company, Meta decided to use books obtained from the LibGen shadow library and later downloaded Anna’s Archive, using torrenting.[1] The plaintiffs argued that Llama can reproduce snippets of their books and that using their works without permission diminished their ability to license them for training.[1]
Decision or outcome
The court rejected both of those arguments. Llama could not generate enough text from the plaintiffs’ books to matter, and the plaintiffs “are not entitled to the market for licensing their works as AI training data”, so whether such a market exists was irrelevant to fair use.[1] The court identified market dilution, meaning that a model can flood the market with competing works, as the potentially winning argument, but the plaintiffs presented no meaningful evidence of it, so the fourth fair use factor favoured Meta.[1] The court noted that Llama is available under a free licence but was developed for commercial reasons.[1] On 2025-06-27 it granted Meta summary judgment on the DMCA claim, because the removal of copyright management information could not further an infringement that was a fair use.[2] A case management conference was set to discuss the separate claim that Meta unlawfully distributed the books during torrenting.[1]
Significance for software licensing and SAM practice
The ruling concerns copyright fair use, not a licence contract, but it bears on how AI training data is licensed. The court said that LLM developers who cannot use copyrighted books without permission would need to license them or use public domain works, and expected large-scale licensing to develop.[1] The decision is limited to the evidence presented about these thirteen authors. For a ruling reaching a different result on AI training, see Thomson Reuters v. Ross Intelligence, and for a parallel authors’ case see Bartz v. Anthropic.
Lessons learned
- The ruling is narrow. The court wrote that it was not holding that training on copyrighted works is lawful, and that plaintiffs with better market-effects evidence will often win.[1]
- Licensing markets were not treated as protected. The lost opportunity to license works for training did not count as market harm under the fair use analysis.[1]
- Licensing attempts leave a record. Meta’s negotiations, the rights obstacles it found and its later decision to use shadow-library copies were all part of the evidence.[1]
- Acquisition and distribution are separate questions. The torrenting distribution claim was not decided by the fair use ruling.[1]