Kadrey v. Meta

The claims

Direct copyright infringement (copying books to train the Llama models); infringement by distribution (alleged reuploading of book data while torrenting); DMCA claims

What has happened

Thirteen authors, including Sarah Silverman and Ta-Nehisi Coates, sued Meta in July 2023 over the books used to train its Llama models. Unsealed filings later showed Meta sourced books from shadow libraries such as LibGen after internal debate, and allegedly redistributed data while torrenting. On June 25, 2025, Judge Vince Chhabria granted Meta summary judgment on fair use for training — while stating the ruling bound only these thirteen plaintiffs and faulting them for failing to build a record of market harm. He wrote that on a better-developed record, copying protected works to train generative models without permission could be unlawful in many circumstances. The separate claim that Meta distributed books to third parties while torrenting was not resolved by that ruling. In July 2026, Chhabria denied the authors’ bid for an immediate appeal, saying the fair-use issues should reach the Ninth Circuit after final judgment as a complete package.

Key developments

• 2023-07-07 — Kadrey, Silverman and Golden file suit in the Northern District of California; ten more authors later join.
• 2023-11-20 — Judge Chhabria dismisses most theories, including the claim that Llama itself is an infringing derivative work, leaving the core copying claim.
• 2025-02-06 — Unsealed filings describe Meta’s internal decision to use LibGen and its torrenting of book datasets.
• 2025-06-25 — Chhabria grants Meta summary judgment on fair use for training, stressing that the plaintiffs failed to develop evidence of market harm and that the ruling is not a blanket blessing of AI training.
• 2026-07-08 — Chhabria denies the authors’ motion to certify an interlocutory appeal of the fair-use ruling; appellate review waits for final judgment.

Why it matters for training data

Kadrey is the fair-use win that reads like a warning. The court sided with Meta because the record was thin, not because training is categorically lawful, and it flagged market dilution as the theory most likely to defeat fair use on better evidence. For data buyers, the open flank is acquisition mechanics: torrenting can create a separate distribution claim even where training itself survives. Suppliers should assume the next plaintiff arrives with the economic evidence this one lacked.

Sources

• Goodwin: N.D. Cal. judge rules Meta’s AI training is fair use
• Perkins Coie: court sides with Meta but leaves door open for future challenges
• Chat GPT Is Eating the World: interlocutory appeal denied (July 8, 2026)
• CourtListener docket: Kadrey v. Meta Platforms

Deeper analysis

• Training-data copyright lawsuits: the landscape

← All case files Weekly lawsuit news