Anthropic pirated half a million books to train Claude.

Anthropic pirated half a million books to train Claude. U.S. District Judge Araceli Martínez-Olguín approved the company’s $1.5 billion class-action settlement on Monday, saying it provides “meaningful relief” for the affected authors. Plaintiff attorney Justin Nelson described the settlement as “the largest known copyright recovery in history.” About 91 percent of the more than 482,000 books covered by the settlement have been claimed by authors or publishers, who are due roughly $3,000 per title. The settlement closes the Bartz litigation, originally filed in 2024 by bestselling thriller novelist Andrea Bartz and two co-plaintiffs. The settlement does not close the question it created.

The instrument. The case proceeded as a Rule 23(b)(3) damages class action, with the settlement requiring court approval under Rule 23(e). The substantive law is the Copyright Act of 1976 (17 U.S.C. §§ 101 et seq.), whose operative provisions — the exclusive-rights grant, the infringement cause of action, the statutory-damages ceiling for willful violations at $150,000 per work, and the fair-use defense — provide the architecture against which the settlement’s price is measured. U.S. District Judge William Alsup, who has since retired, granted preliminary approval in September 2025 from the Northern District of California and produced the doctrinal architecture on which the settlement rests: a bifurcated finding that training AI chatbots on copyrighted books is not itself illegal under the fair-use provision, but that acquiring those books through pirate websites is wrongful under the infringement provision. The split is the load-bearing architecture. It separates two questions the Copyright Act treats as one: whether the use was licensed, and whether the use was fair. The bifurcation’s effect is to make the acquisition method the operative liability trigger, not the training itself.

The steel-man. Anthropic’s strongest defense under this architecture is that training large language models on copyrighted text is transformative, producing a new commercial product from inputs that serve a different purpose than the originals and do not displace the originals’ market. The defense has doctrinal support. The four fair-use factors are fact-intensive, and Alsup’s training-is-fair-use finding rested on record evidence of how Claude’s training process actually uses copyrighted text. Anthropic would also argue that the books were already publicly available on pirate sites at the time of download, that no concrete marketplace harm has been shown, that the $3,000-per-book benchmark represents substantial compensation, and that the 91 percent claim rate demonstrates meaningful class engagement. Each argument has legal traction. The fair-use training finding is the strongest.

The audit. The steel-man does not survive contact with the settlement’s actual arithmetic. The Copyright Act’s willful-infringement ceiling is $150,000 per work. Multiplying that figure by the 482,000 books in the certified class produces an exposure in the tens of billions of dollars — the upper bound the settlement was negotiated against, not the price the settlement actually names. The settlement names $3,000 per book, or roughly two percent of the statutory maximum. The settlement is not the price of infringement. The settlement is the discount off the price of infringement. Alsup’s finding that Anthropic “wrongfully acquired” the books through pirate sites is the legal characterization of willful conduct — the same characterization that triggers the $150,000 ceiling. The settlement names the price for the conduct the court called wrongful. The price is approximately one-fiftieth of the statutory exposure.

The structural significance is what the settlement authorizes. Anthropic’s reported valuation has reached $965 billion. The $1.5 billion settlement is approximately 0.16 percent of that valuation — a figure small enough to be absorbed as a cost of doing business, not large enough to function as a deterrent. The 91 percent claim rate indicates that authors found the recovery meaningful, but the per-book recovery remains a fraction of the statutory maximum and a vanishing fraction of the company’s market capitalization. The settlement does not penalize industrial-scale literary theft. The settlement prices it.

Alsup’s bifurcation — training is fair use, acquisition is not — names the doctrinal framework the settlement establishes. Under that framework, AI companies that license their training data operate inside the fair-use safe harbor. AI companies that download from pirate sites operate inside the infringement zone, with statutory damages as the exposure ceiling. The settlement’s per-book price becomes the litigation discount against that ceiling — a discount available to any future defendant willing to settle before trial. The settlement is what industrial-scale literary theft costs in federal court. The settlement is not what the Copyright Act says it should cost.

Anthropic pirated half a million books to train a commercial product. The federal judiciary priced the theft at roughly one-fiftieth of the statutory maximum. The settlement names the price of theft. The Copyright Act names the price of infringement.