Japanese Books Bought and Destroyed for AI Training Stir Copyright and Preservation Concerns

From Japanese Bookshops to American Scanners

Orders reaching secondhand bookshops across at least 11 regions of Japan from around July 2026 have exposed a growing trade in printed knowledge for artificial intelligence. Buyers sought books on philosophy, history, medicine, law and Edo period culture, sometimes purchasing more than 1,000 volumes from a single shop. Their priority was complete content rather than pristine covers.

The Japanese federation of antiquarian booksellers identified the purchasing activity across those regions. Investigations traced books through Japanese logistics centers to the United States, with destinations including an Amazon facility in Las Vegas and the AI company Anthropic. At scanning operations, book spines were removed, pages were scanned at high resolution and the remaining paper was shredded.

Import and export records put shipments of Japanese books to businesses suspected of involvement at 57.8 metric tons. At an assumed average weight of 500 grams per book, that would represent about 115,600 volumes. The figure is an estimate based on weight; the number sold specifically to Anthropic by a major Japanese distributor remains undisclosed.

The purchases have brought together three disputes: who can authorize the conversion of books into AI training material, whether publishing businesses will lose future sales, and how to protect scarce physical copies. Older specialist books are particularly attractive because they may contain knowledge that has never appeared online. For libraries and researchers, the same scarcity makes their destruction a concern.

A Major Distributor Confirms Sales

Nippan IPS, a subsidiary of Japanese publishing distributor Nippan, entered a contract with Anthropic in November 2024 and sold Japanese books to the company from December 2024 through December 2025. Nippan acknowledged the transactions in a written response dated October 5, 2026, after the Japan Book Publishers Association requested an explanation.

Nippan said it checked the intended use before signing the contract but did not understand the purchases to be for AI training. It also said its policy is to refuse sales when that purpose is explicit. The number of books and the value of the transactions were blacked out in court documents, and Nippan has not disclosed them.

Nippan is a wholesaler, a business that moves books from publishers to retailers. Its involvement places the controversy inside Japan's mainstream publishing supply chain, alongside the purchases from secondhand shops. The confirmed distributor transactions began well before the wave of orders that booksellers noticed in July 2026.

The publishers association sought information about the sales on September 17 and said it would discuss its response at board and council meetings on October 27. The outcome of those meetings has not been described.

How the Purchases Unfolded

The sequence stretches across several years and includes both commercial transactions and a United States copyright case. Keeping those dates together helps explain why the Japanese publishing dispute emerged after Anthropic had already been buying books.

• Early 2024: Anthropic began a book scanning effort known internally as Project Panama.
• November 2024: Nippan IPS signed its contract with Anthropic.
• December 2024 to December 2025: Nippan IPS sold Japanese books to the company.
• June 2025: United States District Judge William Alsup issued a fair use ruling addressing Anthropic's use of books.
• Around July 2026: Large orders from American businesses began reaching secondhand bookshops across Japan.
• July 2026: A United States court gave final approval to a $1.5 billion settlement with authors concerning pirated books.
• September 17, 2026: The Japan Book Publishers Association requested an explanation from Nippan.
• October 5, 2026: Nippan acknowledged the sales in writing.

American AI companies have also collected books from countries other than Japan over several years. Estimates place the broader acquisitions in the millions of volumes, putting the Japanese transactions within a much larger effort to obtain printed material.

Why AI Companies Want Printed Books

A book that has never been digitized offers something an internet search cannot easily retrieve: its full text. Specialist Japanese publications may contain historical records, legal discussion, medical writing or descriptions of cultural practices that are absent from websites.

Training a language model means processing large quantities of text so that software learns patterns in language and information. Books offer sustained arguments, edited prose and explanations that run across many pages. Anthropic's internal material described an interest in learning good writing from books rather than relying on lower quality internet text.

Books published before the widespread adoption of generative AI also offer a large body of writing produced before machine generated text became common. For developers choosing training material, the date and origin of a text can therefore matter alongside its subject.

Removing a spine allows loose pages to pass through scanners efficiently. That industrial advantage creates a direct cost: the physical volume is dismantled. A privately held scan and a book available for consultation in a library serve different purposes. Readers and researchers need access to the text, while historians may also need the original paper, binding, illustrations or handwritten annotations.

What the American Court Decided

Judge William Alsup's June 2025 ruling treated Anthropic's AI training use as fair use and accepted the conversion of purchased printed books into digital copies under the circumstances examined. Fair use is a United States copyright rule that allows some uses without permission, depending on factors such as purpose and effects on the market for the work.

The court treated Anthropic's acquisition and retention of books from pirate sites differently, rejecting fair use for that part of its digital library. The company later reached a $1.5 billion settlement with authors over the pirated material, with final court approval in July 2026.

For readers, the practical point is that the case examined several activities separately: obtaining a book, making a digital copy, keeping it in a library and using its text to train a model. A purchase can affect the legal analysis, while the purpose and circumstances of copying also matter.

Authors and publishing groups argue that paying for a physical book does not itself grant permission to digitize its contents or use them for AI training. Their demand is for developers to obtain authorization from copyright holders before proceeding, rather than treating an ordinary retail transaction as an agreement covering every later use.

Japan Has Its Own Legal and Contract Questions

Japan's Copyright Act contains two provisions central to the debate. Article 26-2, paragraph 2, generally allows a lawfully sold copy to be resold without fresh permission from the copyright holder. That rule helps make ordinary secondhand book trading possible.

Article 30-4 addresses uses that are not intended to enjoy a work's expressed thoughts or feelings, including certain forms of information analysis. AI training can fall within that framework, subject to the provision's conditions and limits, including protection against unreasonable harm to a copyright holder's interests. Selling a copy and analyzing its contents are therefore separate legal questions.

Publishers are also examining their contracts with distributors. Japan's resale price maintenance system allows publishers to set retail prices through agreements within the book trade, under an exception to competition rules. The publishers association asked how large sales to an overseas company outside that contractual network had been handled.

Nippan's conduct is consequently being debated on more than copyright grounds. Publishers want to know the terms on which their books were supplied and whether established commercial arrangements covered those transactions. A court decision in the United States addresses a different legal framework from the Japanese contracts governing distribution.

Rare Copies Face a Different Risk

Japanese scholars have warned that books surviving in small numbers, especially those without electronic editions, could become permanently unavailable if remaining copies are bought and destroyed. The concern is strongest where a volume is difficult to replace and its contents are held only in a private digital collection.

A model trained on a book learns from its text, but readers normally receive generated answers rather than access to a complete, searchable reproduction of that edition. Cultural preservation requires people to be able to inspect records, check passages and revisit the original context. Those tasks depend on accessible copies and reliable cataloging.

The purchases cover both specialist secondhand books and older cultural material. A book about Edo period culture may have been published recently, while another volume may itself be an old artifact. Its age, rarity and availability in other collections determine the preservation risk.

No title by title inventory has been disclosed for the Japanese shipments, leaving the number of rare or irreplaceable volumes unknown. The Japanese Cartoonists Association and publishers have called for copyright holder authorization and government rules that balance technological development with protection of cultural heritage.

Publishing Businesses Are Already Under Pressure

Japan's Research Institute for Publications estimated sales of printed books and magazines at 964.7 billion yen in 2025, down 4.1% from the previous year. The total fell below 1 trillion yen for the first time in about half a century.

Nippan Group's revenue fell from about 521 billion yen in the financial year ending March 2021 to about 382.7 billion yen in the year ending March 2025. That is a decline of 138.3 billion yen, or roughly 26.5%, in four years. The group recorded an ordinary loss of about 1.6 billion yen in the first half of fiscal 2025.

A large order offers a distributor immediate revenue in a shrinking market. Publishers face a longer calculation: authors, editors, proofreaders and designers must be paid, and some titles need repeat printings before their costs are recovered. An AI company may purchase a single copy of a title and use what it learns to answer many users.

Publishers fear that readers seeking practical knowledge will accept an AI summary instead of buying a book. No data has yet quantified how much Japanese book purchasing has been displaced by AI answers. The commercial concern is nevertheless concrete: future demand for books helps finance the next round of writing and publishing.

American distributor Ingram reportedly viewed cutting and scanning books as lawful but was concerned about the effect on its reputation. Its reaction shows why companies may weigh relationships with authors and publishers alongside the immediate value of a sale.

Anthropic's Expansion Adds Urgency in Japan

Anthropic's search for Japanese text comes as its business in Japan grows. At an October 14, 2026, briefing, Anthropic Japan president Hidetoshi Tojo said its workforce had increased from three people to more than 50 in a year, while the number of companies adopting its technology had risen tenfold.

The company's expansion in Japan also coincided with a proposed NTT data center in southern Osaka Prefecture. The estimated construction cost was about 1 trillion yen, with power intake capacity of up to 500 megawatts. The companies were seeking an agreement by the end of fiscal 2026, with partial operations possible in 2028.

Japanese businesses are therefore encountering Anthropic in two roles: as a supplier of AI services and as a customer for material used to develop those services. The book dispute brings the conditions for obtaining that material into the same discussion as investment and adoption.

Publishers are asking for permission and contractual clarity; booksellers have received unusually large orders; scholars want scarce copies preserved. Government action and industry agreements would need to address those separate interests, including how books intended for scanning are identified and how culturally valuable copies remain accessible.

Key Points

• Secondhand bookshops in at least 11 Japanese regions received large orders from around July 2026.
• Shipments weighing 57.8 metric tons would equal about 115,600 books at an assumed 500 grams each.
• Nippan IPS confirmed sales to Anthropic from December 2024 through December 2025, without disclosing quantities or value.
• A United States court accepted fair use arguments concerning purchased books and AI training, while treating pirated library copies differently.
• Japanese publishers are examining copyright, distribution contracts and possible effects on future book sales.
• Scholars warn that destroying scarce copies can put access to historical and specialist knowledge at risk.