Kadrey v. Meta: a court called LLM training on pirated books fair use, but only because the authors failed to prove market dilution

Verification status: IN CHECKING — not publish-ready, not pending, not verified. No green badge is claimed here.

The problem

Where the text used to train a large language model comes from has become the central legal question of the AI era, and Kadrey v. Meta put it squarely before a federal court. Thirteen authors sued Meta over the books used to build its Llama models: the order records that “in this case, thirteen authors - mostly famous fiction writers - have sued Meta for [using their] books to train [its] generative AI models (specifically, its large language models, called Llama)” (source). The plaintiffs include Richard Kadrey, Sarah Silverman and Junot Díaz, a group The Authors Guild describes as “13 well-known and best-selling authors” (source). Their objection was not to Llama’s outputs but to its inputs: the order records that “Meta first used a shadow library” to obtain the books it wanted for training (source).

What was built

The system at issue is Meta’s Llama family of large language models, trained in part on copyrighted books that Meta acquired by downloading them from online “shadow libraries” rather than licensing them; the order names “LibGen, Z-Library, and others” among the sources (source). The court was careful that copying to train and copying to distribute are different legal acts: “reproduction and distribution are separate rights that must be considered separately” under “17 U.S.C. § 106(1), (3)” (source). The parties’ cross-motions the court decided here reached the reproduction/fair-use question, and a Goodwin LLP client alert confirms the frame: “Judge Chhabria of the Northern District of California granted summary judgment for Meta Platforms in an AI copyright infringement suit” (source).

The outcome

Judge Chhabria granted Meta summary judgment on fair use, but the reasoning is the story. On the first factor, the order is unequivocal: “there is no serious question that Meta’s use of the plaintiffs’ books had a ‘further purpose’ and ‘different character’ than the books - that it was highly transformative,” and “this factor favors Meta” (source). The case then turned on the fourth factor, market harm, and specifically on a theory the plaintiffs barely pressed: that Meta had “copied their works to create a product that will likely flood the market with similar works, causing market dilution” (source). The court found the theory potentially decisive but the evidence absent: “the plaintiffs presented no meaningful evidence on market dilution at all. Absent such evidence, the fourth factor can only favor Meta. Therefore, on this record, Meta is entitled to summary judgment on its fair use [defense]” (source).

That is why the win is narrower than its headline. The order opens by warning that the general answer to whether training on copyrighted works without permission is illegal is likely yes: “this case presents the question whether such conduct is illegal. Although the devil is in the details, in most cases the answer will likely be yes” (source). And it closes the same way, disclaiming any broader precedent: the ruling “stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one” (source). The Authors Guild read it the same way, calling it a case where “Meta won only on technical grounds, a matter of procedure, not on the merits of the law” (source). The order was signed by “Vince Chhabria, United States District Judge” and dated “June 25, 2025” (source).

Weakest load-bearing source, named where you meet it: the holding, the “highly transformative” finding, the market-dilution reasoning and the “wrong arguments” language all rest on a Tier-1 primary - the signed summary-judgment order (Document 598), retrieved as the court PDF and text-extracted this session and quoted verbatim above. The softest part of the record is the hosting of that PDF: the copy fetched this session came from a news organisation’s server (Courthouse News), not directly from the court’s own docket. That single-host risk is now cleared: the identical signed order (same caption, “Case 3:23-cv-03417-VC Document 598 Filed 06/25/25”, same “Dated: June 25, 2025” signature block) is independently hosted as a full PDF by Chat GPT Is Eating the World and is docketed as Document 598 by Justia, so the primary no longer rests on one host. The two secondaries used here are a law firm’s client alert (Goodwin, independent of the parties) and The Authors Guild, which is the authors’ own trade body and therefore a party-aligned commentator; its “technical win” characterisation is opinion, and only the verbatim court language above is load-bearing.

How this was verified

• Method: the court order was retrieved as a PDF from the public record and text-extracted this session via a zlib stream decode (the war-room box has no pdftotext). Every quoted holding was matched against that extracted text, and the case caption (“Case 3:23-cv-03417-VC Document 598 Filed 06/25/25”), the signing judge (“Vince Chhabria, United States District Judge”) and the date (“June 25, 2025”) were read from the same PDF. The holding and the market-dilution framing were cross-checked against two independent secondaries (a Goodwin LLP client alert and a statement from The Authors Guild).
• Date verified: research settled 2026-09-06 (maker round 1).
• Cross-host corroboration: the signed order was matched across two independent hosts this session (Courthouse News and the full PDF at Chat GPT Is Eating the World), both carrying the same “Document 598 Filed 06/25/25” caption and the “Dated: June 25, 2025 / Vince Chhabria” signature block, and it is docketed as Document 598 by Justia. The primary does not rest on a single news-site host. The remaining hardening is an archived snapshot against the official government copy (govinfo USCOURTS-cand-3_23-cv-03417). No party-confirmation contact is required or sought; the order is a public, refutable record.

Related case files

• Bartz v. Anthropic - the sister ruling in the same district the same week: where Bartz held training fair use but priced the piracy that fed it at $1.5 billion, Kadrey held training fair use only because the authors never built the market-harm record that could have flipped it.
• Thomson Reuters v. Ross Intelligence - the other side of the AI-training fair-use question, where a Delaware court rejected fair use for training a non-generative legal-research tool, the mirror image of Kadrey’s outcome for a generative LLM.
• Getty Images v. Stability AI - the UK counterpart on training-data copyright, showing how a different legal system reached its own limited answer on whether ingesting protected works to train a model infringes.
• Meta pauses EU AI training on public data - the regulator-driven analogue involving the same company: where Kadrey tested Meta’s book-training data in court, the Irish DPC forced Meta to halt training on user data before it started.

Sources

1. Tier 1 (primary). U.S. District Court for the Northern District of California - Kadrey et al v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, Order on Cross-Motions for Partial Summary Judgment (Document 598), filed June 25, 2025 (Chhabria, J.). Court PDF (hosted by Courthouse News): https://www.courthousenews.com/wp-content/uploads/2025/06/kadrey-et-al-vs-meta-order-motion-partial-summary-judgment.pdf (archive attempt pending - web.archive.org unreachable from the headless war-room this session; local sources/kadrey-v-meta-order-2025-06-25.pdf. Official record: govinfo USCOURTS-cand-3_23-cv-03417.) Second independent host of the identical signed order (same “Document 598 Filed 06/25/25” caption and “Dated: June 25, 2025 / Vince Chhabria” signature block, confirmed by text extraction this session): Chat GPT Is Eating the World, https://chatgptiseatingtheworld.com/wp-content/uploads/2025/06/Judge-Chhabria-Fair-Use-decision-in-Kadrey-v.-Meta-June-25-2025.pdf (local sources/kadrey-v-meta-order-chatgptietw-2025-06-25.pdf); also docketed as Document 598 by Justia, https://law.justia.com/cases/federal/district-courts/california/candce/3:2023cv03417/415175/598/
2. Tier 2 (strong secondary, independent of the parties). Goodwin LLP, “Northern District of California Judge Rules That Meta’s Training of AI Models Is Fair Use,” June 2025. https://www.goodwinlaw.com/en/insights/publications/2025/06/alerts-practices-aiml-northern-district-of-california-judge-rules (archive attempt pending; local sources/goodwin-kadrey-alert.html).
3. Tier 2 (strong secondary, authors’ trade body - party-aligned commentary). The Authors Guild, “Meta AI Ruling: Meta Gets a Technical Win, but the Law Favors Authors,” June 26, 2025. https://authorsguild.org/news/meta-ai-ruling-meta-gets-technical-win-but-law-favors-authors/ (archive attempt pending; local sources/authorsguild-kadrey.html).