AI Policy & Copyright Law

Anthropic's $1.5 Billion Copyright Settlement Is Final

A U.S. federal judge signed off on the largest copyright settlement in American legal history — and the first major AI training-data lawsuit to reach a final, court-approved resolution.

• Anthropic will pay roughly $1.5 billion under the class-action settlement in Bartz v. Anthropic.
• The payout covers about 482,000 titles — worked out to roughly $3,000 to $3,100 per book.
• U.S. District Judge Araceli Martínez-Olguín said the deal gives affected authors and publishers "meaningful relief."

The lawsuit centered on how Anthropic obtained the text used to train its Claude chatbot. A group of authors and publishers accused the company of copying their copyrighted books from pirate libraries — unauthorized collections of millions of titles hosted on file-sharing and torrent sites — rather than from legitimate sources. Training an AI on such material, the plaintiffs argued, was not just a copyright question but a chain-of-custody one: even if the resulting model might qualify as fair use, the books were not lawfully acquired in the first place.

That distinction is the hinge of the whole case. In an earlier 2025 ruling, another judge largely sided with Anthropic on the fair-use question, reasoning that reading and training on copyrighted text is a transformative use similar to how libraries and researchers handle books. But that ruling left the piracy angle untouched. Judge Martínez-Olguín's final approval of the settlement closes that gap: the company can argue its model training was fair use, yet still owe compensation for the illegal route taken to get the training material.

The settlement's scale sets a precedent that every AI lab now watches. A payout of nearly $3,000 per title implies a market price for copyrighted training data that publishers will cite in every future negotiation — and for every other chatbot trained on scraped text. It also reframes "fair use" debates: even if a court accepts training as fair use, the provenance of the data matters. For companies racing to build frontier models, the cheapest dataset may no longer be the safest one.

The case is not without controversy. Judge Martínez-Olguín reportedly expressed reservations about aspects of the deal during approval, and legal observers note the settlement avoids a full trial that might have produced binding precedent on the fair-use question itself. What is settled, however, is the signal: AI companies now face a concrete, nine-figure cost for building their models on unlawfully sourced books.

Knowledge takeaway: The largest U.S. copyright settlement ever (~$1.5B, Bartz v. Anthropic) was approved by Judge Martínez-Olguín. It covers ~482,000 pirated-book titles at roughly $3,000 each. Fair-use defenses for AI training may still hold, but the illegal sourcing of training data is now independently actionable.