Anthropic’s Book-Piracy Settlement Wins Final Court Approval

A federal court has given final approval to what is believed to be the largest copyright settlement in US history, ordering Anthropic to pay $1.5 billion to authors whose books it copied from pirate libraries to train its Claude models. The decision, issued in San Francisco on July 20, 2026, closes the first major US copyright case against an AI developer to end in a payout — and puts a concrete price on training a model with unlicensed data.

U.S. District Judge Araceli Martínez-Olguín granted final approval to the class action, which a group of authors brought over Anthropic’s use of pirated copies of their work without permission or payment. Now-retired Judge William Alsup had preliminarily approved the deal in September 2025 before stepping down, leaving his successor to sign off after a fairness hearing in May 2026. Anthropic, which is backed by Amazon (AMZN ) and Alphabet, admitted no wrongdoing.

What the settlement requires

Under the terms set out by class counsel, Anthropic will pay $1.5 billion plus interest into a fund covering more than 480,000 books — roughly $3,000 per work. The company also has to destroy the original files it pulled from two “shadow libraries,” Library Genesis and Pirate Library Mirror, along with any copies made from them.

For the rest of the industry, the more important detail is how narrow the release is. It covers only how Anthropic acquired the books in the past. It grants no license to keep training on them, and it releases nothing about what Claude actually generates. Authors keep the right to sue over model outputs, and over any book left off the settlement list.

The line the court drew

The payout grew out of a split ruling that has shaped how AI developers read their own legal exposure. In June 2025, Alsup found that training Claude on books was “exceedingly transformative” and therefore fair use — but that downloading and storing more than seven million pirated books in a central library was not.

“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,” Alsup wrote of the training. The piracy was a separate question. Because US copyright law allows damages of up to $150,000 per work for willful infringement, a library that size left Anthropic facing theoretical liability in the hundreds of billions of dollars, with a trial set for December 2025 to fix the figure. The company settled rather than test it.

That distinction — training can be fair use, but sourcing the data from pirate sites is where the liability sits — is the operational takeaway for anyone building a model. It rewards developers who license or lawfully buy their training material and penalizes those who reach for the free, infringing copy. It also sharpens the incentive to scrub questionable data provenance, the kind of IP-washing that has quietly spread through the model-training supply chain.

A template, not a precedent

Anthropic’s is the first of the dozens of AI copyright suits — including cases against OpenAI, Microsoft (MSFT ) and Meta — to settle at this scale, and both sides have treated it as a marker for what comes next. Class counsel Justin Nelson called it “the first of its kind in the AI era” and said it sets a precedent requiring AI companies to pay copyright owners.

Legally, though, a settlement is not a ruling. It binds only the parties and sets no precedent that future courts must follow on whether training on copyrighted work is lawful. What it sets instead is a market price. Defending one of these cases to trial can run into the millions, statutory damages make the downside catastrophic, and so some developers will now license data they might not strictly need — insurance against a bet that could sink a balance sheet. It is a US case, but the warning about American legal risk travels to any lab that trains on English-language books. For authors and publishers, the lesson is that protecting original work now carries real leverage.

What happens next

The response from rightsholders was unusually strong. Claims came in for more than 90% of the eligible works — over 440,000 of them — class counsel told the court, a rate far above the roughly 10% typical of a class action. Anthropic is paying into the fund in four installments through September 2027, with an initial $300 million already sitting in an interest-bearing escrow account; money will not reach authors until the court’s approval survives any appeals.

Not everyone is satisfied. Class counsel had trimmed its fee request to about 12.5% of the fund, below the roughly 30% common in such cases, but objectors still argued the lawyers take too large a cut, that the total is too small to deter a company of Anthropic’s size, and that the deal wrongly shuts out works not registered in the United States. A handful of authors and publishers opted out to pursue their own claims, which remain live. Those suits — and the unsettled question of whether scraping the open web is fair use — are where the copyright fight over AI training moves next.