Used bookshops across Japan are fielding unprecedented bulk orders from unfamiliar buyers, amid suspicion that overseas AI companies are buying up old books to scan for training data, Daily Shincho reported.
The shift began in July, when accounts believed to belong to a proxy buyer started placing large orders several times a day. The titles ranged from art books to specialist histories, with no pattern to what was ordered. Staff at one shop began asking each other what was going on.
"We do get customers who buy in similar ways for resale, but the scale of these orders was far too large," said Yoshihiro Sumida, head of Jimbocho Yomitaya, a used bookshop with a store in Kichijoji, western Tokyo, and an online sales site. "There is no doubt that orders in a form we had never seen before are increasing."
The shop normally sells about 1,000 books a month online. In August and September that figure rose by roughly 500. Sales have since calmed somewhat, Sumida said, but the unusual bulk orders have not stopped.
The pattern has spread to shops nationwide, and the trade has buzzed with talk that foreign AI companies are buying up Japan's old books. In the United States, lawsuits brought by authors revealed that the major AI company Anthropic had bought millions of used books for training. The company was also confirmed to have cut the books up and destroyed them in the end.
The buyers placing the Japanese orders are not a single company, according to Sumida. Several firms are involved, and their names are similar and their shipping addresses close together, suggesting they sit under one group. The books were sent to several locations in western Japan.
"What we cannot tell is whether these orders really come from AI companies in the United States or elsewhere," he said.
Why old paper is worth buying
Hidenori Kawamura, a professor at Hokkaido University's graduate school of information science and technology, said the appeal is straightforward. Information only began to be digitized in earnest about 20 years ago, so much of the knowledge accumulated before that exists only on paper.
For AI companies, breadth matters more than the quality of any single item, Kawamura said. Training on a vast body of data puts individual errors into perspective, so even inaccurate material can be absorbed.
Anthropic once contacted Kawamura's laboratory for permission to use a paper he co-authored as training data. The paper was not especially famous, but the company evidently judged that specialist material of that kind still held information worth learning from.
IT journalist Hiroshi Mikami said cost explains the method. Apart from a few expensive titles, used books are cheap, so buying them up is more cost-effective than signing a comprehensive deal with publishers to scan everything from new releases to back issues.
The result, he argued, is that profits from the publishing industry are being siphoned off without payment. Japan has not only a large body of work by domestic authors but also translations of literature from around the world, read in Japanese, he said.
"That rich publishing culture could be bought up cheaply by overseas companies, have the best parts extracted, and be weakened," Mikami said. "From the standpoint of protecting Japanese culture, this deserves more debate."
Countermeasures will not be easy to put in place. Daily Shincho's fuller treatment of the story, carried on its Shinsho QUE service, examines who the bulk buyers are and whether Japan's publishing culture can be protected from AI companies.