Anthropic copyright settlement with authors wins court approval

A federal judge approved Anthropic PBC’s $1.5 billion settlement with authors and publishers over books used in building Claude, putting a dollar figure on one of the AI sector’s most closely watched copyright fights. The deal matters because it separates two issues that AI labs would prefer to treat as one: whether model training can be fair use, and whether acquiring copyrighted works from pirate libraries creates separate liability.

U.S. District Judge Araceli Martínez-Olguín called the settlement fair and adequate, writing that the payment gives class members meaningful compensation given the unsettled legal theories at stake. She also noted that the authors and publishers faced litigation risk, since a trial loss could have left them with nothing.

The settlement covers roughly 500,000 works that Anthropic copied from pirate libraries while developing Claude, according to the plaintiffs. Authors and publishers are set to receive $3,000 for each covered work. Anthropic must also destroy the pirated materials.

Judge William Alsup had issued preliminary approval of the settlement last year before retiring. His earlier ruling gave both sides something to claim: he found that training AI systems on copyrighted books qualified as fair use, but also held that downloading those books from piracy sites was not protected and infringed copyright.

Training fight remains unresolved

The case does not settle the broader question of whether AI companies can use copyrighted material to train models without a license. Alsup’s fair-use finding remains a notable win for Anthropic, but the settlement turns on how the company obtained the books, not on a final industry-wide rule for model training.

The plaintiffs alleged that Anthropic used copyrighted books sourced from illegal websites including Library Genesis and Pirate Library Mirror. In court filings, they accused the company of trying to profit from the creative work embodied in those books, arguing that its model was built on human expression taken without authorization.

For AI companies, the order sharpens the compliance problem around training data provenance. The court’s treatment suggests that a fair-use argument may not protect a lab if the underlying dataset was assembled from pirated copies. That distinction is likely to matter for companies that collected large corpora before licensing became a board-level issue.

Anthropic deputy general counsel Aparna Sridhar said after the judgment that the company reached the settlement in 2025 following the court’s ruling that training AI on books is fair use under copyright law. She said more than 91% of covered authors and publishers have claimed their share of the payment and that Anthropic expects to close the matter.

The settlement lands while other AI copyright cases continue against major technology companies and model providers. Companies including Google LLC, Meta Platforms Inc., Midjourney Inc., Perplexity AI Inc. and OpenAI Group PBC are still facing claims tied to copyrighted works and AI systems, with each looking for rulings that preserve the economics of large-scale training.

The $1.5 billion payout, described as the largest copyright class-action settlement in history, is large enough to change how AI companies assess data risk. It does not create a licensing template for the sector, and it does not answer every fair-use question. It does show that courts may treat scraped or copied training data very differently depending on how it was obtained.

This story draws on original reporting from SiliconANGLE.