A federal judge has approved Anthropic’s $1.5 billion settlement with authors and publishers whose pirated books were collected while the company was building the technology behind Claude.

The money is enormous, but the more interesting part is the legal line quietly forming between teaching an AI and stealing the textbooks.

U.S. District Judge Araceli Martínez-Olguín granted final approval to the class-action settlement on July 20, making it the largest known payout in a U.S. copyright case.

The agreement resolves claims from authors who accused Anthropic of downloading unauthorized copies of their books and storing them in a massive digital library.

The lawsuit began in 2024 with authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, but eventually expanded to cover hundreds of thousands of books and the writers, publishers, and other rightsholders connected to them. More than 91 percent of the authors and publishers covered by the deal submitted claims for compensation, according to Anthropic.

Under the settlement, eligible rightsholders were originally expected to receive roughly $3,000 for each qualifying book, although the final amount can depend on ownership arrangements, duplicate claims, administrative costs, and how the funds are divided. The official settlement website says the deadline to file a claim was March 30, 2026, while the deadlines to object or leave the settlement have also passed.

This case is slightly more complicated than the familiar story of an AI company feeding copyrighted material into a machine and hoping nobody notices. In June 2025, Judge William Alsup ruled that using books to train an AI model could qualify as fair use because the process was transformative. Claude was not simply storing novels and handing them back to users like a suspiciously efficient bookstore photocopier. The model was analyzing language patterns to create new responses.

But that decision came with a very large asterisk.

The court found that Anthropic had collected more than 7 million pirated books and stored them in what was described as a “central library,” including material obtained from Books3, Library Genesis, and Pirate Library Mirror. The judge drew a distinction between what the company eventually did with the books and how it obtained them in the first place. Training an AI may be fair use in some circumstances. Downloading millions of pirated files is still downloading millions of pirated files.

That distinction is likely to become one of the most important ideas in the next phase of AI copyright law.

Tech companies may be able to argue that model training transforms source material enough to qualify as fair use, but that does not necessarily give them permission to build their datasets through shadow libraries and illegal downloads.

The algorithm does not erase the paper trail.

Anthropic faced a trial that could have exposed the company to damages reaching into the hundreds of billions of dollars. Instead, the two sides reached the $1.5 billion agreement in 2025, removing a potentially existential legal risk for one of the largest companies in generative AI.

The company also agreed to destroy the original pirated files covered by the dispute. Court records previously showed that Anthropic later moved toward more conventional acquisition methods, including purchasing physical books, removing their bindings, scanning the pages, and using the resulting digital copies for research and training. It is a deeply industrial solution to a digital copyright problem, but it illustrates the legal difference between buying a book and grabbing it from an unauthorized archive.

Not everyone believes the settlement goes far enough. Some authors argued that the payment was too small compared with the scale and value of Anthropic, while others objected to attorney fees or said the agreement excluded certain copyright owners. Martínez-Olguín rejected those objections, saying complaints about the settlement amount did not realistically account for the risks of taking the case to trial. The court awarded the plaintiffs’ attorneys more than $101 million in fees, considerably less than the $187.5 million they requested.

A number of writers and publishers opted out and are continuing with separate lawsuits, which means the broader fight is nowhere near finished. More than 25 writers, including Dave Eggers and Vendela Vida, previously filed another complaint after leaving the settlement. Similar cases involving OpenAI, Microsoft, Meta, Apple, Midjourney, and other companies are also testing how copyright rules apply when enormous collections of human work become raw material for AI systems.

For Anthropic, the ruling closes one major chapter while preserving the most valuable part of the earlier decision: the finding that training an AI model on copyrighted books can be fair use.

For authors, the settlement establishes that the method used to assemble a training library still matters, and that piracy can carry consequences substantial enough to show up even on a major AI company balance sheet.

The outcome is neither a complete victory for creators nor a blank check for AI developers. It is something more practical and probably more influential: an early legal blueprint. Companies may be allowed to teach machines using copyrighted culture, but the courts are making it increasingly clear that they still need to pay attention to how that culture gets through the door.