Kadrey v. Meta

What the authors allege

Thirteen authors, including Sarah Silverman and Ta-Nehisi Coates, sued Meta in July 2023 over the books used to train its Llama models: direct copyright infringement for copying the books to train the models, infringement by distribution over the alleged reuploading of book data while torrenting, and related DMCA claims. The suit was among the earliest and most closely watched of the author-led AI training cases, given Meta’s scale and the prominence of the named plaintiffs.

In November 2023, Judge Chhabria dismissed most of the plaintiffs’ theories — including the claim that Llama itself is an infringing derivative work — leaving the core copying claim and the separate torrenting-distribution claim to proceed. That early narrowing set up the case to turn almost entirely on fair use.

What unsealed filings showed

Filings unsealed in February 2025 described Meta’s internal decision-making around sourcing books from LibGen, a shadow library, despite internal debate over the choice, and described Meta’s use of torrenting to acquire book datasets — a method that, by its peer-to-peer design, can also redistribute data to other users.

That detail matters for the case’s remaining distribution claim, because it undercuts any suggestion that Meta was unaware of the provenance and legal risk of its sourcing choices when it made them. It also gave plaintiffs their strongest evidentiary hook for a claim that survived independently of the fair-use fight over training itself.

The ruling and its limits: fair use, but not a blank check

On June 25, 2025, Chhabria granted Meta summary judgment on fair use for training. He was explicit that the ruling bound only the thirteen plaintiffs in this case, and that they had failed to develop evidence of market harm — the factor he identified as the one most likely to defeat a fair-use defense on a stronger record.

Crucially, the ruling did not resolve the separate claim that Meta distributed books to third parties while torrenting them. That claim survives independently of the fair-use win on training, keeping a live theory of liability open even after Meta’s central defense succeeded.

“On this record, Meta’s use of the plaintiffs’ books to train its Llama models was fair use — but on a better-developed record, including evidence of market harm, such copying could be unlawful in many circumstances.”— Order Granting Summary Judgment, No. 23-cv-03417 (N.D. Cal. June 25, 2025)

Where it stands: no early appeal

On July 8, 2026, Chhabria denied the authors’ motion to certify an interlocutory appeal of the fair-use ruling, holding that the fair-use issues should go up to the Ninth Circuit as one package after final judgment rather than piecemeal.

That means resolution of the still-open torrenting and distribution claim will likely need to happen first, or alongside any further proceedings, before appellate review of the fair-use holding can occur — extending the case’s overall timeline well past the June 2025 summary-judgment ruling that made headlines at the time.

Why it matters for training-data buyers

Kadrey reads like a fair-use win with a warning attached. The court sided with Meta because the record on market harm was thin, not because training on copyrighted books is categorically lawful — and it specifically flagged market dilution as the theory most likely to defeat fair use with better evidence.

For data buyers, the open flank is acquisition mechanics, not just content selection: torrenting can create a separate distribution claim even where the underlying training use survives a fair-use defense. Anyone licensing or acquiring training data should assume the next plaintiff arrives with the market-harm evidence this record lacked, and should scrutinize how a dataset was physically acquired, not only what it contains.

That distinction — training use versus acquisition method — is likely to shape how future plaintiffs structure their claims, splitting a single dataset’s legal exposure into two separately litigable questions rather than one.