Anthropic's pirated-book training ruled illegal; legally bought books cleared as fair use

Judge finds Anthropic trained Claude on pirated books

A federal judge has ruled that Anthropic illegally obtained copyrighted books to train its Claude models, in a decision arising from Bartz v. Anthropic, a lawsuit filed in May 2026. According to The Stanford Daily, the ruling found that Anthropic used pirated copies of works by more than ten Stanford professors, including Tobias Wolff, Abraham Verghese, Chang-Rae Lee, Daniel Mason, Hayagreeva Rao and Myra Strober.

The case ended in a $1.5 billion settlement paid to authors and publishers of nearly 500,000 registered works, which the Authors Alliance describes as the largest copyright settlement ever involving an AI company.

A split ruling on piracy versus fair use

Judge William Alsup's decision drew a sharp line between how training data is obtained and whether training itself is lawful. He barred Anthropic from pirating books to build a permanent central library, but held that training Claude's large language models on legally acquired books constitutes fair use.

Alsup compared machine training to a human reader absorbing a book's themes and style. "Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different," he wrote.

Authors opt out and sue again

Not every plaintiff accepted the settlement. Wolff, a professor emeritus of English and National Medal of Arts winner, and Chang-Rae Lee opted out of the Bartz deal and in May 2026 filed a separate non-class action, Cruz v. Anthropic, seeking a jury trial and full statutory damages of up to $150,000 per infringed work. The Bartz settlement offered roughly $3,000 per work.

The Cruz plaintiffs allege Anthropic illegally downloaded, reproduced and distributed copyrighted books from "shadow libraries" such as LibGen and PiLiMi, and that the company broke the law by storing the pirated files in a central library. Ten of Wolff's publications were used in training without his permission.

"To put me to work that way, without remittance, is a form of slave labor," Wolff told The Stanford Daily, adding that the settlement "still gives Anthropic pretty free rein to roam through our literature." If the Cruz plaintiffs succeed, they could win a permanent injunction barring Claude training on their works.

The licensing problem remains unsolved

Paul Goldstein, a Stanford Law scholar specializing in intellectual property and himself a plaintiff in the class action, framed the core difficulty as one of transaction costs: millions of individual copyrighted works may be copied in the course of a single training run, making it impracticable for a platform to license every work, or even any single one.

Goldstein also noted that most industrialized countries grant authors a moral right of attribution, ensuring credit for their work, while the United States does not — a gap many authors care about beyond payment.

Broader litigation and a push for legislation

The Anthropic cases sit within a much larger wave of legal action. More than 70 infringement lawsuits have been filed against AI companies by authors, visual artists and news organizations, including The New York Times and Getty Images.

The Authors Guild is pursuing its own cases, including one against OpenAI and Microsoft, as part of what Chief Executive Officer Mary Rasenberger described as a multi-pronged effort to make clear that AI companies must license the books they use. The Guild is also working with Congress on national licensing legislation. In an email to The Stanford Daily, Rasenberger said licensing would let authors and publishers require filters ensuring their works do not appear in model outputs unless separately licensed.

Why it matters

The ruling sets a significant precedent: acquiring training data through piracy is indefensible, while training on legally purchased books is fair use. That split gives AI companies a clear compliance path — buy the books — and a clear liability trap for shortcutting it, with the $1.5 billion settlement quantifying the cost. The pending Cruz case could push per-work damages fifty times higher than the class settlement, and a permanent injunction against training on specific works would be an even stronger remedy. Yet Goldstein's transaction-cost point shows the harder problem is unresolved: licensing millions of works at scale has no working mechanism, and the legality of mass scraping more broadly — the question behind dozens of other lawsuits — remains uncertain.