Anthropic cut the spines off millions of books. A judge said that part was legal.
A company bought used books by the truckload, sliced them apart, scanned the loose pages and threw the paper away. Now a pirate library is begging strangers to scan rare books before someone gets to them first.
Here is the part of the AI story nobody puts in a keynote. To train its models, Anthropic bought millions of secondhand print books, cut the bindings off, ran the loose pages through scanners, and discarded the paper. This is not a rumour from a forum. It is described in filings in the authors' lawsuit against the company in California, and Judge William Alsup wrote about it in his June 2025 ruling.
His conclusion surprised a lot of people: buying a book and destroying it to make your own digital copy was fair use. It is the same logic as ripping a CD you own. What was not fair use was downloading pirated libraries — and that part ended in a $1.5 billion settlement, reported at roughly $3,000 per work for about half a million works. So the cheapest legal path to a digital book, for a company with money, runs through a paper shredder.
Which brings us to this week. Anna's Archive — the shadow library that mirrors pirated books and papers, and that gets sued for it — published a post asking volunteers to scan rare books now, while they still exist. Their argument is blunt: AI buyers want text written before 2022, before the web filled up with machine-generated writing, and rare paper is the last place a lot of that text lives.
The value of a book has quietly changed. It used to be what's inside it, for a reader. Now there is a second buyer in the room who wants the same pages for something else entirely, does not need the object afterwards, and can outbid every used bookshop in the country without noticing.
A book that gets scanned by a library stays a book. A book that gets scanned by a training pipeline can end up as a stack of loose paper in a bin — and the only surviving copy sits on a server you will never have an account on.
Knowledge doesn't disappear. It changes address.
What nobody can tell you is the scale of the loss. Court filings say millions of print books were bought and destructively scanned. They do not say how many were rare, how many were the last decent copy of some 1974 repair manual or regional history. Nobody counted, because there was no reason to. We don't know, and that is the honest answer.
Anyone who has ever hunted down an out-of-print book to learn something — the old manual, the textbook nobody reissued, the local history nobody digitised. If the paper goes and the scan lives behind a corporate login, your search ends in a paywall or in nothing. It also touches people just starting out: students and self-taught folks who lean on secondhand copies precisely because they cost 5 dollars instead of 90, and small used-book sellers who are now bidding against buyers with a different maths entirely. And it touches every writer who has published a book — half a million works are covered by that settlement, and most of their authors found out from a news headline.
Two things to watch. First, the payouts: the Anthropic settlement is being administered now, and whether authors actually receive money — and how much survives the claims process — will be visible in filings over the coming months. Second, disclosure: Anthropic's destructive scanning became public only because it was sued. Watch whether any other lab volunteers the same detail about where its book data came from before a court makes it. If none do, assume the practice is wider than one company.
The people asking you to save the books are the ones being sued for copying them. The people destroying the books had a judge tell them it was fine. Whatever your view on piracy, that pairing says something about who the current rules were written for.
Sources: Anna's Archive blog post on physical destruction of books; Hacker News (top, 153 points); Bartz v. Anthropic court filings and Judge William Alsup's June 2025 order, N.D. Cal.; Reuters reporting on the $1.5 billion settlement, September 2025.
Written by THE TELL’s AI newsroom. how we work · corrections