What happened

  • Thirteen authors, including Richard Kadrey, Sarah Silverman and Ta-Nehisi Coates, sued Meta in the Northern District of California (Case No. 3:23-cv-03417-VC), alleging Meta copied their books without permission to train its large language models.
  • The plaintiffs moved for partial summary judgment and Meta cross-moved (Dkt. 482, 501). On June 25, 2025 the court denied the plaintiffs' motion and granted Meta's on fair use as to training (Dkt. 598).
  • Two days later the court granted Meta summary judgment on the plaintiffs' DMCA claim, reasoning that because the copying was fair use, removing copyright management information could not have furthered an infringement (Dkt. 601).
  • The June 25 order did not decide whether Meta distributed the books while torrenting them. It set a case management conference for July 11, 2025 to discuss that claim.
  • By the December 12, 2025 Joint Case Management Statement (Dkt. 661), the parties described the torrenting and distribution claim as the sole remaining issue. Plaintiffs had moved on December 11, 2025 for leave to file a Fourth Amended Consolidated Complaint that would limit the class to registered works, add contributory infringement and add three loan-out companies as plaintiffs. Meta said it would oppose.
  • Meta said it would assert fair use against the torrenting claim. Plaintiffs said Meta had not previously raised a fair use basis for that part of the case. The statement records that the court's schedule was extended after new evidence of torrenting from Anna's Archive and Library Genesis (Dkt. 641) and that Meta proposed a further extension of about ten weeks (Dkt. 660).

What the court said

The court was explicit that its ruling is narrow. It wrote that the decision "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful" and that it applies only to these plaintiffs on this record. The order's conclusion says no earlier case answers the fair use question for LLM training, and that in cases "like Meta's, it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects."

Why Meta won anyway: the court treated Meta's use as highly transformative, so the plaintiffs needed to win decisively on market harm. It found they "presented no meaningful evidence on market dilution at all." The court said market dilution, meaning a flood of competing AI-generated works, is "so important in this context" that evidence of it could have gone to a jury or even won the motion, and that the plaintiffs offered none.

The court also criticized the contrary reasoning in Bartz v. Anthropic (N.D. Cal., June 23, 2025), calling its analogy between training models and teaching schoolchildren "inapt."

What it means for your firm

  • Our reading: this is a ruling about a thin record, not a safe harbor. A client in publishing, media or software that trains or buys AI models should read it as "market dilution evidence decides the case," not "training is fair use."
  • Provenance is the live exposure. The surviving claim is about how the books were acquired and shared, not how they were used to train. Ask AI vendors where their training data came from, not just whether it was fair use.
  • A split between judges in the same district (this order and Bartz) means no settled rule in the Northern District of California. Vendor contracts that rely on a favorable rule should say which ruling they assume.
  • The docket is still moving. A brief like this one is only as current as the last filing read, so check the live docket before you cite it.

What to do Monday

  1. Add a data provenance question to your AI vendor diligence: where was the training data sourced, and did the vendor acquire any of it from shadow libraries or torrent sites.
  2. Check any indemnity in your AI vendor agreements for whether it covers copyright claims about training data and about distribution of that data, since the open claim in this case is about distribution.
  3. For clients who publish text, record what evidence they could show on market harm from AI-generated substitutes, since that evidence decided this ruling.
  4. Pull the live docket (Case No. 3:23-cv-03417-VC) before citing this case and note the date you checked.

Sources

Questions and answers

Did Meta win Kadrey v. Meta?

Partly. On June 25, 2025 the court granted Meta summary judgment on fair use for training Llama on these thirteen authors' books, and on June 27 it granted summary judgment on the DMCA claim. A claim about torrenting and distributing the books was still pending in the latest filing we read, dated December 12, 2025.

Does the ruling mean training AI on copyrighted books is legal?

No. The court said the opposite is likely in most cases, and said its ruling turned on the plaintiffs' failure to present evidence of market harm. It applies to these plaintiffs on this record.

What is the torrenting claim?

Plaintiffs allege Meta downloaded books from shadow libraries using BitTorrent and in doing so made them available to others. Meta denies distributing the plaintiffs' works and says it will assert fair use. The court had not ruled on it in the filings we read.

Operational information, not legal advice. Facts come from the filings on the court docket linked under Sources, read on Oct 27, 2026. A docket moves: this brief covers only the filings we read and does not replace the live docket. The reading of what it means for your firm is ours.