Book publishers sue Meta over Llama’s word-for-word piracy

Type two sentences from a college calculus textbook into Llama, and Meta’s AI will happily recite the rest of the section word-for-word.

To five major book publishers, that is not impressive engineering. It is intellectual property theft on a historic scale. Macmillan, McGraw Hill, Elsevier, Hachette, and Cengage, alongside author Scott Turow, have filed a massive class action lawsuit against Meta, accusing the tech giant of committing one of the most massive copyright infringements in history.

Tapping pirate hubs to feed open-source AI

The publishers pulled no punches in their filing. They allege that Meta knowingly scraped copyrighted books and academic papers directly from notorious pirate websites, including LibGen, Anna’s Archive, Sci-Hub, and Sci-Mag, just to pump them into the Llama models.

The lawsuit also takes aim at Meta’s use of the Common Crawl dataset, alleging it is already packed with pirated works.

According to the complaint, the consequence of scraping pirate libraries is obvious. Llama frequently delivers verbatim and near-verbatim substitutes of original texts. The publishers demonstrated this using James Stewart’s Calculus: Early Transcendentals, showing that feeding the AI two brief introductory sentences prompted it to immediately reproduce the upcoming section word-for-word.

Internal leaks and the $1.5B Anthropic precedent

Meta has found itself in this corner before. Earlier lawsuits from authors unearthed awkward internal discussions where Meta staff debated how to handle media coverage suggesting the company was actively using datasets it knew were pirated.

A federal judge did rule in Meta’s favor in an earlier copyright dispute last year. Still, that same judge explicitly noted the decision did not establish that training language models on copyrighted works was lawful.

The shadow hanging over Meta right now is Anthropic. When authors hit Anthropic with a class action over pirated training data, a judge ruled that training on legally acquired books qualified as fair use, but greenlit claims regarding millions of pirated files. Anthropic chose to settle that suit last year for a staggering $1.5 billion.

Turow and the publishers are demanding monetary damages, an injunction to halt the practice, and an order forcing Meta to disclose every copyrighted book and article used to train Llama. Meta spokesperson Dave Arnold pushed back, stating that courts have recognized training on copyrighted data can qualify as fair use, and promised to fight the suit aggressively.

Why it matters

This case could shatter the tech industry’s favorite shield: the fair use doctrine.

AI companies have long argued that scraping public web text is transformative because models only learn patterns. But that argument falls apart the moment a model starts regurgitating textbook sections verbatim, and it collapses completely if the data was pulled straight from known pirate websites.

If the court forces Meta to disclose its full training inventory, it will blow the lid off the open-weights AI ecosystem. Developers relying on Llama derivatives could suddenly face thorny legal and compliance risks if downstream models are officially branded products of infringement.

In our view, Meta’s biggest vulnerability is not just paying a penalty, but the precedent Anthropic already set. Paying a billion-dollar settlement is bad enough, but having to scrub training pipelines and pay royalties for every academic paper and textbook will fundamentally change the economics of building frontier models.