TOKYO — A book that exists in perhaps a dozen known copies documents the ceremonial textile practices of Edo-period merchant houses in Osaka. It was compiled around 1810 and hand-illustrated. In August 2026, the owner of a Tokyo used-book platform received a bulk order that included it — along with several hundred other titles — from an account that had placed similar orders with sixteen other stores in the same week. The shipment was consolidated and sent to a warehouse in Okayama Prefecture. What happened to the book after that is unknown.
Japan’s used-book market has reported an anomalous surge in bulk buying since roughly August 2026, with individual sellers describing days when sales ran close to five times their normal volume. The orders share a pattern: they come through multiple accounts, cover an eclectic mix of categories — philosophy, medicine, political history, law, and materials from the Edo period spanning 1603 to 1868 — and are consolidated at a single logistics facility. Export records subsequently surfaced by Japanese media show a group company affiliated with a major Japanese publishing distributor shipped more than 50 tons of goods labeled “JAPANESE BOOKS” to the United States since the preceding year, a volume that translates to roughly 100,000 volumes at a standard paperback weight.
No company has confirmed placing those orders. No entity has identified the Okayama warehouse as its own. The books are simply gone.
The working assumption inside Japan’s antiquarian trade is that the volumes are bound for AI training data pipelines. That assumption has a precedent the industry knows well. Publishers and authors have spent the past two years pressing cases against every major AI laboratory in US courts, with results that range from massive settlements to rulings that leave the practice fundamentally intact.
The clearest template is the one Anthropic built. Beginning in early 2024, the company launched an internal program described in its own documents as “our effort to destructively scan all of the books in the world.” The project, run internally under the name Project Panama, was led by Tom Turvey, the former head of partnerships at Google Books. The acquisition method was industrial: large orders placed with booksellers, books delivered to facilities equipped with hydraulic cutting machines that removed the spines, industrial flatbed scanners that captured each page, and disposal of the physical originals once digitized. The goal was access to text that existed only in print — older material, out-of-print academic works, and pre-digital categories conspicuously absent from commercial databases.
Anthropic’s operation eventually generated the largest copyright settlement in recorded history. Authors Andrea Bartz, Charles Graebner, and Kirk Wallace Johnson led a class action in the Northern California District Court, resulting in a $1.5 billion agreement covering roughly 500,000 eligible works. But Judge William Alsup’s ruling distinguished between Anthropic’s earlier use of pirated copies, which was indefensible, and Project Panama itself. The physical purchase and destruction of legally acquired books, the judge ruled, constituted fair use under US law.
That ruling is now the legal floor on which any American company operating a similar program can stand. Physically purchase the books, legally. Scan them. Discard them. Under current US copyright law, the authors’ rights end when the book is sold.
The Tokyo Antiquarian Booksellers’ Cooperative Association has confirmed it is aware of the unusual purchasing activity. Speculation within the trade, the organization said, has centered on AI training. Rare and irreplaceable editions — books that may exist in only a handful of copies, documenting Edo-period customs, pre-Meiji medicine, or early Meiji political philosophy — are not replaceable if they leave circulation permanently.
The scale of what Claude AI and its competitors require to continue improving is difficult to compress into a single figure. Claude now leads twenty-six percent of Anthropic’s internal research and development. The models that power that kind of autonomous research capacity are trained on data that exists only because someone, somewhere, physically acquired the books that contained it, cut off their spines, and ran them through a scanner.
What makes the Japan story distinct from the Project Panama revelations is the cultural specificity of the material and the absence of any legal accountability. When Anthropic scanned English-language works in the United States, American authors could sue in American courts and did. The unauthorized data access cases working through Australian courts right now involve AI agents accessing digital systems without permission — systems with server logs and traceable credentials. Physical books shipped from Okayama to a US warehouse leave different evidence, subject to different law, in a country whose citizens are not parties to the transaction.
Anthropic has not confirmed or denied involvement in the Japanese purchases. As 404 Media’s investigation found, and as Euronews documented in its account of Project Panama, the companies acquiring books to scan have generally preferred not to discuss where the books come from or what happens to them afterward.
The hydraulic cutting machine at the center of Project Panama does not distinguish between an American paperback and an Edo-period manuscript printed on washi paper. The legal system that authorized its American operation has no jurisdiction in Okayama.
