Anthropic copyright settlement: judge approves $US1.5bn deal

A federal judge in California has approved Anthropic’s $US1.5 billion settlement with authors and publishers who alleged the company trained AI models on pirated books. The order, made by Judge Araceli Martínez-Olguín in the Northern District of California on 20 July, gives the AI sector a rare hard number for copyright exposure tied to training data.

The court order covers 506,194 works. Compensation is set at $US3,000 a work, or about $4,600, for a total fund of $US1.5 billion, about $2.3 billion. Ars Technica reported that 350 authors opted out. Anthropic said more than 91 per cent of covered authors and publishers had already claimed payment.

That is why the approval matters outside this one docket. Boards, investors and publishers no longer have only a theoretical fair-use fight to price. They have a settlement figure accepted by a federal judge, even if the underlying legal questions remain unsettled. The approved fund also gives lawyers and finance teams something plainer to discuss with model vendors: what copyrighted training material might cost when a case ends in settlement rather than judgment.

Martínez-Olguín said the settlement would deliver “meaningful relief” to the class. Her order does not decide whether AI developers can lawfully train models on books in other circumstances. It resolves this case on these terms, which makes the outcome narrower than a precedent but more concrete than the policy arguments that have surrounded generative AI since the first book-training suits were filed. That distinction is important for rivals watching their own copyright cases move through court.

Lead plaintiffs said the approval should send a warning beyond Anthropic, according to Ars Technica’s report.

“brings us one step closer to real accountability for Anthropic and puts all AI companies on notice they can’t shortcut the law or override creators’ rights.”

Lead plaintiffs, via Ars Technica

Anthropic, best known for its Claude models, used a more contained formulation. Aparna Sridhar, the company’s deputy general counsel, told Ars Technica that Anthropic was pleased more than 91 per cent of covered authors and publishers had claimed payment and was looking forward to closing the matter. The company is trying to treat the dispute as a booked legal cost, not a live fight over how Claude was built.

The opt-out number helps explain why the deal survived approval. Class settlements can become fragile when a large group of claimants decides the terms are too narrow or the payments too small. Here, 350 opt-outs sat against more than half a million covered works. For most eligible parties, certainty appears to have beaten the prospect of a longer copyright case. The authors who opted out may still press their own claims, but they are a small share of the class described in the order.

The wider AI market should not read the order as a template for every training-data lawsuit. Image claims, book claims and contract disputes all involve different evidence, and a settlement is not a final ruling on whether training is lawful. Still, Anthropic’s deal gives courts, publishers and model developers a figure to cite as copyright claims move from policy debate to balance-sheet risk. It also adds pressure on AI companies to document where training material came from before a dispute reaches that point.