Anthropic Destroyed Millions of Books to Train Claude: Was That Legal?

Anthropic bought millions of used books, sliced off the covers, and scanned every page to train Claude. A federal judge ruled that destroying books this way was legal.

Court files exposed the program in January under the internal name Project Panama. Brokers now sell the same book-buying service to other artificial intelligence (AI) firms.

Why Anthropic Destroying Books Was Ruled Legal

The rule is older than AI. When you buy a book, it is yours. You can resell it, lend it, or bin it. Lawyers call that the first-sale doctrine.

Judge William Alsup applied it plainly. Anthropic “purchased its print copies fair and square,” he wrote in his June 2025 order. So the authors were not arguing about payment. They were arguing about the switch from paper to PDF.

Alsup found that switch added no copy at all.

“Here, every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy. The print original was destroyed. One replaced the other,” wrote Alsup in the Bartz v. Anthropic case.

Here is the strange part. Destroying the book is what kept it legal. Keep the paper and you hold two copies. Shred it and you hold one.

Mass book scanning has reached court before, and the precedents cut both ways.

Inside Project Panama, Anthropic’s Book Scanning Program

Anthropic tried the free route first. It had “many places from which” it could buy books, Alsup wrote. Instead, chief executive Dario Amodei took them, to skip the “legal/practice/business slog.”

Cofounder Ben Mann downloaded Books3 in early 2021, a library of 196,640 pirated titles. Five million more came from Library Genesis that June. Two million followed from the Pirate Library Mirror in 2022.

Buying started later. Anthropic hired Tom Turvey in February 2024, who had run book deals at Google. His brief was to obtain “all the books in the world.” Contractors then cut the bindings, trimmed the pages, and scanned each copy.

The order names no supplier. However, the Washington Post reported two, the used-book sellers Better World Books and World of Books. It also described hydraulic cutters removing the spines.

Only the piracy cost money. A judge approved a $1.5 billion payout on July 20, worth roughly $3,000 per book. Anthropic must delete the pirated files within 30 days of judgment. It still faces a separate $75 million claim and music publishers suing Anthropic over lyrics.

Why Old Books Became the Hottest AI Commodity

Old print is clean, and that is the whole point. AI text now floods the web. Feed it back into a model and the model degrades. A 2024 Nature study called this model collapse.

So books printed before ChatGPT cannot contain machine writing. Brokers spotted the gap. ISBNdb sells book sourcing to AI firms. “The world’s best AI training data is sitting on a shelf,” its site says.

Meanwhile, the cost barely registers. Anthropic raised $65 billion in May at a $965 billion valuation, on revenue running at $47 billion a year. The settlement equals about 3% of that.

For crypto readers the thread is provenance. Proving where data came from is now a legal cost rather than a bonus. That same problem drives decentralized AI market forecasts and the case for blockchain based data ownership.

Not everyone accepts the trade-off. Elon Musk, chief executive of SpaceX and Tesla, says his team will scan differently.

Follow us on X to get the latest news as it happens

Gentle scanning has existed for decades. It is simply slower and dearer. So the open question is whether public anger moves an industry the law currently rewards for shredding.

The post Anthropic Destroyed Millions of Books to Train Claude: Was That Legal? appeared first on BeInCrypto.