AI companies need enormous amounts of written material to train increasingly capable models, and much of that data has traditionally come from websites and other digital sources. However, physical books have also joined the pile, becoming valuable because they contain specialised knowledge and writing that may not be easily available online.
This growing demand is now drawing attention to Japan’s used-book market, where bookstores have reported unusually large purchases since around August.
Some sellers have seen sales rise to about five times their normal levels, while hundreds of books across subjects including philosophy, history, medicine, and law are being bought in bulk. As such, this unusual pattern has fuelled suspicions that at least some of the books are being used for AI training.
Why These Book Orders Stand Out
The size of the orders is only one reason booksellers have taken notice, as the books being purchased do not follow the usual pattern of demand in the secondhand market.
According to reports based on an investigation by Nippon TV, buyers have been purchasing specialised books covering subjects such as Edo-period culture and Japanese politics from decades ago alongside philosophy, medicine, history, and law. One seller reported selling more than 1,000 books, while another noticed orders arriving at regular intervals and suspected that an automated system was placing them.
Multiple buyer accounts were also reportedly sending their purchases to the same logistics centre in Okayama Prefecture. A bookstore in western Japan also separately received an enquiry from an overseas company asking whether it could supply tens of thousands of books.
And this trail extends beyond Japan. Export records examined by Nippon TV showed that more than 50 tonnes of Japanese books had been shipped to the United States since 2025 by a company connected to a major Japanese book distributor. At an estimated 500 grams per book, that amount could represent about 100,000 books.
AI Companies Have Bought Physical Books Before
The suspicion surrounding the Japanese purchases is also linked to evidence that AI companies have previously bought physical books on a huge scale.
Court documents from a copyright case against Anthropic revealed that the Claude developer spent millions of dollars buying millions of printed books to train its models. Contractors removed the bindings, cut the pages, scanned the books into digital files, and discarded the physical copies. Anthropic then kept the digital versions in a research library, from which books were selected for training its AI models.
And this case has particularly gained attention in Japan. On September 17, the Japan Book Publishers Association asked major distributor Nippan to explain reports that it had sold large quantities of Japanese books to Anthropic. However, Nippan has not publicly confirmed or denied individual transactions.
Japan’s AI Copyright Rules Raise More Questions
Japan’s copyright rules also make this issue more complicated. Article 30-4 of the country’s Copyright Act allows copyrighted works to be used without permission in some cases where the purpose is data analysis or AI training and the use does not involve enjoying the work in its original expressive form. However, the exemption has limits, including cases where the use unfairly harms the interests of copyright holders.
For booksellers, the effect of this buying spree is higher sales, but the bigger question remains what happens to their books afterwards and their copyrights.
The situation also shows how far AI companies may be willing to go in search of material that can keep improving their models to fuel the AI race. Training data has become one of the resources that keeps the AI industry moving, and as companies exhaust or compete for material already available online, physical books offer another source of human knowledge, ideas, research, and writing they want to take from.
