How Anthropic ended up paying $1.5 bn for training Claude on pirated books

A US court has approved Anthropic's $1.5 billion settlement with authors over pirated books used to train Claude, while reinforcing that training AI on legally acquired books can qualify as fair use

A US federal judge has given final approval to Anthropic's landmark $1.5 billion settlement with authors who accused the artificial intelligence (AI) company of using pirated books to train its AI assistant and large language model Claude. 

The settlement, approved by a judge in San Francisco, comes as AI developers including Anthropic, OpenAI and Meta face a growing number of lawsuits from authors, publishers, news organisations and other copyright owners over the use of copyrighted material to train large language models.

 

Anthropic said more than 91 per cent of eligible authors and publishers have already claimed their share of the payout.

 

 

"We reached this settlement in 2025, after the court's landmark ruling that training AI on books is fair use under copyright law, which remains the law today," Anthropic deputy general counsel Aparna Sridhar said in a statement quoted by Reuters. "We are pleased that more than 91 per cent of authors and publishers covered by the settlement have claimed their share of the payment, and we're looking forward to bringing this matter to a close."

What was the lawsuit about?

The lawsuit was filed in 2024 by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, who alleged that Anthropic copied millions of books without permission while developing Claude.

 

Court filings showed that the company built a vast digital library using two different sources. It purchased millions of physical books that were later digitised. Separately, it downloaded more than seven million books from online piracy repositories, including Library Genesis (LibGen) and Pirate Library Mirror (PiLiMi).

What did the court decide?

US District Judge William Alsup drew a clear distinction between books Anthropic had legally purchased and books it had obtained through piracy.

 

He ruled that training AI models using legally acquired books amounted to fair use under US copyright law because the court found that the books were not being copied to replace the originals, but were being used to teach the AI model how language works. That made the use "exceedingly transformative" under US copyright law. He also held that digitising legally purchased print books into digital format for internal AI training was permissible, treating it as a format conversion rather than unlawful copying.

 

However, regarding pirated books, he ruled that downloading and maintaining a permanent internal library built from illegally obtained books constituted copyright infringement, even if those books were later used for AI training.

 

The ruling left Anthropic facing potentially enormous statutory damages if the case went to trial. Instead, the company agreed to create a settlement fund of at least $1.5 billion, with eligible authors and publishers expected to receive around $3,000 for each qualifying work before legal fees and other costs are deducted.

Why has 'Project Panama' attracted so much attention?

Although the settlement centred on pirated books, the litigation also revealed how Anthropic sourced legally purchased physical books for AI training.

 

Previously unsealed court documents showed that, under an internal initiative known as Project Panama, the company spent tens of millions of dollars buying used books from booksellers and libraries before removing their bindings with industrial cutting equipment, scanning every page using high-speed production scanners and then recycling or disposing of the physical copies.

 

Internal planning documents, first reported by The Washington Post, described the initiative as an effort to "destructively scan all the books in the world" and instructed employees not to discuss the project publicly. Within a year, millions of books were acquired and disposed.

 

However, Judge Alsup concluded that it was lawful because Anthropic had legally purchased the books before converting them into digital copies for internal use.

Is destructive book scanning becoming an industry trend?

According to a report by 404 Media, an independent technology publication that covers digital rights and internet culture, companies supplying books to AI developers have increasingly begun sourcing physical books in bulk, including rare and out-of-print editions, for destructive scanning.

 

Among the companies serving this market is ISBNdb, a commercial book database that has historically provided bibliographic information to publishers, booksellers and libraries but has increasingly positioned itself as a supplier of physical books for AI training.

 

According to the report, ISBNdb argues that books published before the widespread adoption of generative AI provide cleaner training data because they are free from AI-generated content. The company facilitates bulk purchases of up to one million books for AI training and acknowledges that destroying books on such a large scale could trigger public backlash.

 

Booksellers in Europe and North America have also reported unusually large purchases of specialised academic titles, foreign-language books and rare editions, raising concerns that valuable physical copies may permanently disappear from circulation as AI companies race to secure training material.

What does the settlement mean for the AI industry?

The settlement brings Anthropic's dispute over these pirated books to an end, but it does not resolve the wider debate over whether AI companies can use copyrighted material to train their models.

 

Some authors and publishers opted out of the settlement and continue to pursue separate lawsuits against Anthropic. The agreement also applies only to past claims involving books included in the class list and does not prevent future litigation over new copyright issues or alleged infringement arising from Claude's outputs.

 

The ruling also provides an early roadmap for how courts may view AI training and copyright. It suggests that judges are more willing to accept AI models being trained on legally obtained material than on datasets built from pirated books.

 

As lawsuits against Anthropic, OpenAI, Meta and other AI developers continue, the case is expected to serve as one of the most influential early precedents in defining how copyrighted works can be used to train artificial intelligence systems.