Anthropic's internal program, codenamed "Project Panama" in a court exhibit, for destructively scanning used books to train its AI models, pays premium prices for pre 2022 books and destroys the only copy after training on it.
Inside Barter Books in Northumberland, Stuart Manley recently processed an order that would normally have taken him a week. A Canadian buyer placed a single bulk request of two or three thousand books and cleared it in one transaction. "I've never seen the like of this after 30 years in the second-hand book trade," Manley told the BBC. Around 300 miles south, David Tobin at north London's Walden Books describes sales as booming. In Edinburgh, Derek Walker of McNaughtan's says the pattern is unmistakable.
Anthropic is using a process internally called "Project Panama" to feed its models on used books, and the process is designed to leave no copy behind. The company wants to scan each book once, train on it, and then destroy the only physical copy so the resulting model cannot be accused of redistributing copyrighted text. A court exhibit surfaced in the class action Bartz v. Anthropic PBC lays out the rationale: an internal memo recommends a codename "because we don't want it to be known that we are working on this," and a summary of the project, as the Guardian read it, is to "destructively scan all the books in the world."
Anthropic needs physical books because digital libraries of the right vintage are exhausted. A scanned PDF of a 1970s paperback is often worse than the paper original, and any book published after roughly 2022 risks being contaminated with text that a previous model already wrote. So a buyer walks into a used-book shop, pays a premium, ships the stock to a scanning facility, and the facility destroys the paper after a single pass. The bookseller's good week is the model's training corpus.
The buyer in Manley's anecdote was a Canadian company, not Anthropic, and the booksellers suspect the order is for AI training data. The lab's name is rarely at the till. That separation is part of why a courtroom exhibit, not a bookseller's invoice, is what surfaces the practice.
Tobin told the BBC it is "very nice to sell some of these titles which haven't been sold for many years," and added: "it would be sad if they are ultimately destroyed." Manley is not naming the buyer in print, and Walker is watching the same pattern from the Edinburgh side. Their worry is not theoretical: the copy on their shelf is the only copy that will exist after the scan.
A court has already held that scanning books to train a language model is fair use, and Anthropic's destructive twist is the part that closes off a piracy claim. If there is no surviving digital file, no one is distributing the underlying text. The model has only learned statistical patterns from it. The structure forced the parties in Bartz v. Anthropic PBC to settle. Gizmodo reported the settlement at $1.5 billion, which works out to a few thousand dollars per work claimed in the suit. The exhibit the Guardian read is part of the court docket for that case, and the internal Project Panama memo is Attachment 21 to Document 554.
Anything published before the broad adoption of consumer LLMs in late 2022 carries a premium because it is reliably human-written. Out-of-print academic press runs, the kind that sit in a stack at Walden until they finally move, are now a feedstock. That is why a bookseller in Northumberland can clear a week in a single afternoon. The pre-2022 cut-off is starting to reshape what "rare" means: a 1980s literary fiction title that used to sit unsold is now a piece of training data, while a 2024 literary fiction title, by definition, might not be. The same bookseller is pricing the same shelf against two very different buyers.
The bulk orders hitting UK shops are concentrated in the vendors who are doing this work at all, and many used-book sales still go to human readers. The category claim, that a chunk of recent bulk-buying is AI training data, sits beside the counterpoint that not every shelf is being emptied for a shredder. Anthropic is the named example, not the only buyer. The Project Panama exhibit names the project, but it does not name a price-per-book, a list of titles, or a contract with a specific UK bookseller.
The Canadian buyer in Manley's anecdote has not been identified in court records. The settlement, reported at $1.5 billion, is being processed. And a pre-2022 hardcover that used to sit for years on a shop shelf is now a piece of a model that may never see the light of a reading lamp again.