Anthropic pays $1.5 billion for destroyed books: when AI learns by stealing

Anthropic was forced to pay publishers and authors $1.5 billion after buying and destroying tens of thousands of books to train its artificial intelligence. A case that changes everything.

There is a precise moment when the future of artificial intelligence stopped being purely a technological question and became a legal case worth one and a half billion dollars. That moment involves books: ones that Anthropic bought, cut apart, scanned at high speed, and then sent to recycling in order to train Claude, one of the most sophisticated language models in the world.

To understand what really happened, you have to start at the beginning. AI systems like Claude, ChatGPT, or Meta's models can write, reason, and answer complex questions. But to learn how to do that, they needed something to practice on: millions of texts. The internet offers material in abundance, but books are a different category entirely. The language is richer, more structured, more precise. As early as 2023, a co-founder of Anthropic had publicly stated that books teach artificial intelligence to write well, rather than imitating the poor quality language that floods the web.

The problem is that those books have authors, and therefore also copyright. Asking permission from millions of writers is a slow, expensive, and often impossible process. So the big tech companies looked for shortcuts.

The first was the crudest: downloading pirated books. Anthropic, Meta, and OpenAI drew from illegal digital libraries, collections of millions of volumes shared online without any authorization. Meta did so using torrent connections, being careful not to leave a traceable trail. According to documents that emerged during the trial, this practice was reportedly approved directly by Mark Zuckerberg, despite internal misgivings. An engineer at the company wrote in an internal communication: "Torrenting from the company laptop doesn't feel right to me." And yet they pressed on regardless.

The second shortcut, the one adopted by Anthropic under the so-called Project Panama, was more sophisticated but no less controversial. Instead of downloading pirated copies, the company physically purchased books in batches of tens of thousands, cut them with industrial machinery, scanned them at high speed, and then sent them to recycling. Purchased legally, in other words, without violating copyright at the point of acquisition. But destroyed in the process.

In June 2025, an American court ruled that using books to train artificial intelligence is legal, as it constitutes a transformative process: the AI does not reproduce the texts, it uses them to learn, much like a student who reads in order to grow. But downloading pirated copies is a different matter, and on that point the ruling was unequivocal. Anthropic was required to pay one and a half billion dollars to publishers and authors. Each writer involved can receive up to three thousand dollars per title.

The case does not affect Anthropic alone. Google, Microsoft, and OpenAI are all involved in similar lawsuits, still open or in the process of being resolved. A law professor at Cornell University summed up the situation with a phrase likely to outlast the verdict itself: "AI companies convinced themselves that the rules of academic research applied equally to a billion-dollar business. A colossal mistake."

What this story ultimately reveals is that the way artificial intelligence learned to speak with us carries a cost that someone, until now, chose not to pay.