Japan's Used Bookshops Are Being Bought Out by the Ton

Something odd is happening in Japan's second-hand book trade. Since around August, online used-book sellers have reported a run of unusually large orders, some describing days when sales hit roughly five times their normal volume. The books are being bought by the ton, and the strongest suspicion, still unproven, is that they are headed for AI training.

The pattern is what caught booksellers' attention. According to reporting by Tom's Hardware and Cybernews, the orders arrive through many separate buyer accounts spread across different shops, then converge on a single logistics centre in Okayama Prefecture. From there, at least one shipment of more than fifty tons was sent to the United States. The buying is not random. It leans heavily on philosophy, history, political history, medicine, law and accounts of daily life in the Edo period, the kind of dense, specialist text that is scarce online.

The word attached to the practice is "destructive scanning": strip the binding, run the pages through a scanner, then discard the originals. It is fast and it is final. The physical book does not survive the process, which is why rare-book watchers are alarmed that first editions and out-of-print volumes could be feeding a data pipeline and then vanishing.

Here a note of caution is owed. No public evidence ties these particular Japanese purchases to any named company. Destructive scanning entered the wider conversation through separate United States litigation, in which Anthropic, the maker of this site's namesake, was reported to have bought and destructively scanned print books to build a training corpus. That is a different matter from the Okayama shipments, and treating the two as one story would be a mistake. What links them is a method, not a proven buyer.

The law adds another wrinkle. Japan's Copyright Act includes Article 30-4, an unusually permissive clause that allows copyrighted works to be used for "information analysis" without the rightsholder's consent. It is one reason the country has been floated as friendly ground for training data. But the same article withdraws that permission where a use would "unreasonably prejudice the interests of the copyright holder," a limit that has never been tested against warehouse-scale buying and shredding.

For now this sits between fact and rumour. The surge in orders is real and documented. The destination and purpose are inferred. What is not in doubt is the appetite. Whoever is buying, they want text that a crawler cannot reach, and they want a great deal of it. That demand has already changed how the industry treats web data and licensed catalogues. Paper was supposed to be the safe medium. Apparently not.