Anthropic заплатит $1,5 млрд авторам книг: рекордное урегулирование и победа AI-лабораторий

Anthropic must pay $1.5 billion to book authors — the largest copyright settlement in the history of U.S. class actions. The money was awarded not for training AI, but for downloading about 482,460 works from pirate databases.

Why the record turned into a win for AI labs

The record payout punishes Anthropic for the way it obtained the books, not for the fact that it trained a model on them. The company pulled works from pirate libraries — and it is paying precisely for that violation, while the court upheld the principle that "training AI on legally purchased texts is allowed." For the entire industry, this is a key signal: training on copyrighted content is not equated with theft if the data was obtained legally. That is precisely why a $1.5 billion loss simultaneously became a strategic win for AI labs.

• Settlement amount — $1.5 billion, a record for a copyright class action
• Basis — downloading about 482,460 works from pirate databases
• On average that works out to roughly $3,100 per book
• The case was heard by Judge William Alsup, Northern District of California
• The court ruled that training AI on legally acquired books is fair use

What exactly Anthropic is paying for

Anthropic is paying for piracy, not for training its models. As The Decoder notes, the $1.5 billion compensates authors for the downloading of about 482,460 of their books from illegal databases — essentially for the fact that the texts were obtained unlawfully, not for how they were later used. Had the company purchased the same books legally, no legal claims against the model's training would have arisen, according to the court's logic. The distinction is subtle, but it is exactly what determines who is responsible, and for what.

What Judge Alsup ruled

In June 2025, Judge William Alsup ruled that training AI on legally obtained books falls under the fair use doctrine. He called such use "transformative" — that is, creating a fundamentally new product rather than copying the original. This distinction is what turned a record-breaking loss on paper into a win in substance: Anthropic was punished for the source of the data, but not for the training technology itself.

"The use of the books to train the models was quintessentially transformative," — from

Judge William Alsup's ruling.

Why the sum became a record

$1.5 billion is the largest copyright settlement in the entire history of U.S. class actions. The record is explained by simple arithmetic: nearly 482,460 works, multiplied by compensation of roughly $3,000 each, add up to a billion-and-a-half total. At the same time, under U.S. copyright law, statutory damages for a single willful infringement can reach up to $150,000 per work — so under a worse-case scenario the payout could have been several times larger, and the settlement became a way for Anthropic to limit its risk.

What this means

The precedent sets the rules of the game for the entire industry: it is fine to train models on copyrighted content, but the data must be obtained legally. By paying $1.5 billion, AI labs got an expensive but unambiguous benchmark — pay for your sources, rather than fear lawsuits over the mere fact of training models.

FAQ

What is Anthropic paying $1.5 billion for — training AI?

No. The payment covers the downloading of about 482,460 books from pirate databases, that is, the illegal way the texts were obtained. The court did not find the training of the model on this data to be a separate violation.

Did the court rule that training AI on books is legal?

Yes, provided the source is legal. Judge William Alsup ruled that training AI on legally acquired books is "transformative" use and falls under the fair use doctrine.