Meta Trained Llama on Pirated Books — Trial Now Set for 2027

By Daniel Rao, TechPulse USA

Meta spent years arguing that AI training data is just... out there. This week, the people who own the books decided to test that theory in the courtroom where it hurts most.

Five of the world's biggest publishers — Elsevier, Cengage, Hachette, Macmillan and McGraw Hill — plus bestselling author Scott Turow have quietly moved their class-action copyright war against Meta into its decisive arena. On Thursday, they filed a motion asking that their case, freshly transferred to San Francisco, be handed to the one judge who has already shown he will not let Meta have fair use for free. Responses are due tomorrow. If the judge agrees, Meta's biggest copyright fight lands on the same docket as its first trial, now set for May 3, 2027.

This is no longer a skirmish. It is the publishers' industry, with its lawyers and its catalogues, coming for the way Llama was built.

Photo: The Phillip Burton Federal Building and US Courthouse in San Francisco, where the publishers' class action against Meta now sits after its transfer from New York. Credit: Sam Wheeler, CC BY-SA 3.0, via Wikimedia Commons.

The case that just moved coast to coast

Let's be honest about the timeline, because this story has one. The publishers sued Meta and Mark Zuckerberg personally back on May 5, 2026, in federal court in Manhattan — a 65-page complaint that reads less like a legal document and more like an accusation of grand theft. We read the full complaint ourselves (SDNY Case 1:26-cv-03689), and the language is not subtle.

Then, last month, the case changed shape. Judge P. Kevin Castel granted Meta's motion to transfer the suit from the Southern District of New York to the Northern District of California, where Meta is headquartered. The publishers' case — now N.D. Cal. No. 4:26-cv-11415 — landed in the same courthouse as Kadrey v. Meta, the authors' suit over the same Llama training data.

And this week, the publishers made their real move. On October 16, they filed an administrative motion asking that their case be formally assigned to Judge Vince Chhabria as a case "related" to Kadrey. Their argument, per the docket: both suits concern Meta's training of Llama on downloaded collections of pirated books and journal articles and its torrenting, and keeping them before one judge avoids duplicated work and conflicting results. Here is the part that should worry Meta: the motion states that Meta itself agrees the cases are related.

Responses are due October 19. If granted — and with both sides in agreement, refusal would be unusual — the publishers' class action joins the cluster of Meta cases Judge Chhabria is managing. The same Judge Chhabria who told Meta's lead counsel, flatly, "this case is going to trial on May 3," moving the first Meta book-piracy trial up to May 3, 2027.

What the publishers say Meta actually did

The complaint's core allegation is a procurement story. The publishers claim Meta did not stumble across their books on the open web. They claim it went shopping at the black market.

"They first illegally torrented millions of copyrighted books and journal articles from notorious pirate sites," the complaint states, "and downloaded unauthorized web scrapes of virtually the entire internet. They then copied those stolen fruits many times over to train Meta's multi-billion-dollar generative AI system called Llama. In doing so, Defendants engaged in one of the most massive infringements of copyrighted materials in history."

The "notorious pirate sites" are named: Library Genesis (LibGen), Anna's Archive, Sci-Hub and Sci-Mag. The suit also targets the Common Crawl dataset Meta used, alleging it is "full of unauthorized copies of copyrighted works." And it accuses Meta of stripping copyright management information from the works — a separate federal violation that has survived in several of these AI cases even where the copyright claims themselves fell.

The complaint names Zuckerberg as a defendant personally, alleging he "personally authorized and actively encouraged" the scheme. It even reaches for his company's old religion: "Defendants Meta and Zuckerberg followed their well-known motto: 'move fast and break things.'" The plaintiffs further allege that Meta explored licensing publisher content for training in early 2023 — internal teams discussed a dataset licensing budget as high as $200 million — but abandoned the talks after the issue was escalated to Zuckerberg, according to reports on the filing.

The human texture is in the exhibits. Meta's own engineers, the complaint alleges, worried about the optics: one employee warned that "[i]f there is media coverage suggesting we have used a dataset we know to be pirated," it would be damaging. Another set of discussions weighed how to handle the legal exposure of torrenting. They torrented anyway, the publishers say.

Photo: The Main Reading Room of the Library of Congress — the world of published books and journals whose owners say Meta looted their catalogues to train Llama. Credit: Carol M. Highsmith, public domain, via Wikimedia Commons.

The calculus textbook test

This is the part of the complaint that will be hardest for Meta to laugh off. The publishers did not just allege copying in the abstract. They prompted Llama with two brief sentences from Cengage's bestselling textbook, Calculus: Early Transcendentals, 9th edition, by James Stewart — and, they say, "Llama begins reproducing word-for-word the continuation of the section." The complaint prints a side-by-side chart of the textbook text and Llama's output, with red text marking verbatim copying and blue marking near-verbatim copying, paragraph after paragraph of a freshman calculus chapter.

The sample works named in the suit read like a bestseller list and a syllabus at once: N.K. Jemisin's The Fifth Season, Peter Brown's The Wild Robot, Scott Turow's novels, James Stewart's calculus, plus millions of journal articles from Elsevier's 2,500-plus journals. The authors behind the five publishers' catalogues, the Associated Press noted when the suit was filed, include Turow, James Patterson, Donna Tartt, former President Joe Biden and at least two Pulitzer Prize winners.

Turow, the bestselling author of Presumed Innocent, did not mince words. "I find it distressing and infuriating that one of the top-10 richest corporations in the world knowingly used pirated copies of my books, and thousands of other authors, to train Llama, which can and has produced competing material, including works supposedly in my style," he told The New York Times.

That last phrase — works "supposedly in my style" — is the publishers' real fear in one sentence. Not that Llama memorised a textbook. That it can now be the textbook, or the novelist, on demand, without the author seeing a cent.

Meta's answer: fair use, aggressively

Meta has vowed to "fight this lawsuit aggressively," telling the AP in a statement: "AI is powering transformative innovations, productivity and creativity for individuals and companies, and courts have rightly found that training AI on copyrighted material can qualify as fair use."

And here is the wrinkle the publishers are counting on. "Courts have rightly found" is doing a lot of work in that sentence — because the ruling Meta is leaning on came with a caveat sharp enough to cut. In the Kadrey case last year, Judge Chhabria did rule for Meta on the fair-use question. But he pointedly added that his ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful."

Translation: Meta won the round, not the war. And now the publishers' case — with its piracy allegations, its torrenting evidence, its Zuckerberg-personally narrative — is walking into Chhabria's courtroom, where the judge has already shown he takes market dilution seriously and where, notably, the copyright-management-information claims survived to trial. Chhabria even denied the Kadrey authors' bid for an interlocutory appeal on the downloading question, keeping the pressure on. Meanwhile, across the country, the Third Circuit's ROSS v. Thomson Reuters ruling rejected AI's fair-use defence outright — the first US appeals-court ruling on AI training and fair use, with a footnote warning generative-AI builders specifically.

There is also the Anthropic precedent hanging over everything. Anthropic — which, court documents revealed, physically bought and dismembered hundreds of thousands of books in an operation dubbed "Project Panama" — agreed to pay $1.5 billion to settle its authors' class action. That number is now the industry's rough price for getting the training-data question wrong.

Why this suit could matter more than the rest

There have been many AI copyright suits. Authors sued. Newspapers sued. But publishers are different animals. Individual authors have limited litigation budgets and scattered catalogues; five publishers bring institutional legal teams, deep pockets, and millions of works under one roof. Elsevier alone publishes more than 2,500 journals and 39,000 books. When that scale of plaintiff asks a court for discovery into how Llama was trained, the paper trail gets very long, very fast — and very expensive to defend.

And the timing could not be more pointed. Meta is simultaneously fighting on another front we covered this week: it quietly deleted the AI-training clause from its new business terms, even as legal experts warned the scraping authorisation survives the edit. Read the two stories together and a pattern emerges: a company trying to lock down the future's consent while the past's liability compounds.

Whether Meta can keep claiming fair use will decide not just this case but how every AI lab buys its next dataset. If the publishers prove Meta knowingly sourced training data from pirate libraries — not the open web, not licensed archives, but LibGen and Anna's Archive — then "fair use" stops being a training-data doctrine and becomes a piracy defence. Courts have been strikingly consistent on one point across this whole wave of litigation: they will forgive a lot, but they draw a hard line at piracy.

One more signal of how crowded this fight is getting: last week, a separate group of authors — including New York Times reporter John Carreyrou — filed their own copyright suit naming Google, OpenAI, Anthropic, Meta and Perplexity, and, for the first time in any AI-training case, Elon Musk's xAI, according to Gadgets 360. The AI copyright war now has more fronts than any one company can staff.

Watch October 19. If the publishers' case lands before Chhabria, May 3, 2027 stops being someone else's trial date. It becomes Meta's.

Sources: Associated Press (Hillel Italie, May 5, 2026); the publishers' 65-page class-action complaint (Elsevier et al. v. Meta Platforms, Inc., S.D.N.Y. Case 1:26-cv-03689, filed May 5, 2026); PayAuthors News docket reporting (Oct 14–16, 2026); The New York Times; Variety (via Moneycontrol, May 6, 2026); WebProNews; Gadgets 360 (Oct 16, 2026).

About the author: Daniel Rao covers AI law, copyright and the business of technology for TechPulse USA. He reads court filings so you don't have to.