Anthropic to pay $1.5 billion in book piracy settlement

Anthropic will pay $1.5 billion to settle copyright claims from book authors after a federal court in San Francisco approved a class action deal over books the company downloaded from LibGen and PiLiMi. The order matters for AI companies because the payment addresses allegedly pirated copies, while a prior ruling in the case left standing a favorable fair-use finding for training on books obtained legally.

According to the court materials posted by Justia, Anthropic downloaded books from the piracy databases between 2021 and 2022. The settlement covers about 482,460 listed works, of which 91.3% were claimed. That works out to about $3,000 per claimed work, roughly four times the statutory minimum cited in the case materials.

The settlement also requires Anthropic to destroy the pirated files. It does not resolve every possible copyright theory authors may pursue against the company. Under the deal, authors keep claims tied to AI outputs that reproduce protected works and claims over Anthropic’s future conduct.

The size of the payment is unusual for copyright litigation. The materials describe it as the largest copyright settlement in class action history. Anthropic did not avoid a large cash cost, but the structure of the case gives AI developers something they have been seeking in litigation: a distinction between how training data is acquired and whether training itself is lawful under fair use.

The fair-use ruling remains the industry signal

Judge Alsup previously ruled that training AI systems on legally acquired books was fair use, calling it “transformative - spectacularly so.” That holding is the part AI labs will focus on. The $1.5 billion settlement punishes the use of files from piracy repositories, rather than declaring that model training on books is unlawful as a category.

That distinction is now central to the legal exposure facing AI developers. Many frontier model companies relied heavily on large text corpora gathered from the open web, licensed datasets, archives and other sources. The Anthropic case suggests that courts may treat lawful acquisition and downstream training differently, which could narrow the most damaging claims if a company can show it had a legitimate copy of the material.

The case leaves a large unresolved issue for the industry. Whether mass collection of internet content without permission from authors or site owners counts as lawful acquisition remains open. That question is likely to matter more than the settlement amount for companies that trained models on broad web crawls rather than on explicitly pirated databases.

For authors, the settlement is a large recovery and preserves some claims, including over outputs that allegedly reproduce their works. For AI labs, the outcome is more mixed than the headline number suggests: Anthropic pays heavily for piracy-linked conduct, while the court’s fair-use reasoning on legally obtained books remains available to defendants in other AI copyright fights.

This story draws on original reporting from The Decoder.