The New Yorker Details Anthropic's Book Scanning Project, Revealing More Insights

The New Yorker supplements details about Anthropic's book scanning project disclosed by the court: the internal project, Project Panama, is described as "destructive scanning of all books in the world," where books are disbound, pages cut, scanned, and the physical copies discarded, sending the text into Claude's training.

Earlier this year, U.S. second-hand booksellers received unusual orders through channels like Alibris: a Brooklyn bookseller, who typically sells fewer than ten books a year, suddenly received an order for over ten books at once, specifically targeting obscure titles, averaging about $75 to $80 each, with buyers named Green Parrot Project and Red Sparrow Project, among others, as limited liability companies. A bookseller placed tracking cards in the books, which were sent to ARC Document Solutions' industrial scanning facility in Addison, Illinois; the company advertises that some centers operate around the clock, with a single client processing volumes equivalent to "the size of a football field, stacked four boxes high" each month. The magazine collected over 600 books suspected to be sold to AI companies, mostly from the 1970s to 2010s, with many first editions having only about a thousand copies; eight out of the ten most common publishers are academic presses, with about one-third of the sample being academic publications, including law, history, literary criticism, and a 1972 manual on municipal sewage sludge utilization. Researcher Mycal Tucker testified that non-fiction is generally more valuable for training than fiction, enhancing performance in fields like philosophy and astronomy.

Initially, bulk purchases could be made from wholesalers, but later shifted to second-hand websites to supplement the long tail missing from the internet and existing databases. Anthropic claims it does not acquire antique or collectible rare editions, distinguishing between "rare and valuable" and hard-to-find professional reference books. Related copyright litigation documents previously disclosed that the scale of procurement and scanning is measured in millions of copies, with the company seeking suppliers to process hundreds of thousands to two million copies within about six months. Other reports indicate that the settlement amount in the case is around $1.5 billion, and the judge previously ruled on whether the internal use of digital copies constitutes fair use.

From a market mechanism perspective, this is a replenishment of the paper long tail against web corpus. Beneficiaries are second-hand obscure books and industrial scanning facilities; the pressured parties are the extremely low print academic inventories and publishers that must pay copyright for training data. Funds flow from model companies' procurement budgets into booksellers and scanning outsourcing, with physical copies exiting circulation after scanning.

The 600 books are a sample, not the entire list. The tracking cards prove the logistics endpoint but do not automatically prove that each order entered the same model.

Source: Public Information