Used bookstores in Japan have a strange new problem. Buyers are ordering old books by the box, then by the ton. Many people believe these are books for AI training: they get shipped abroad, scanned into computers, and then shredded.
It sounds like a conspiracy theory. Part of it is proven, and part of it is still a guess. Here is what we actually know, why AI companies want old paper books so badly, and whether any of it is legal.
What is happening to Japan’s used books?
Starting around August 2026, online used bookstores in Japan saw a sudden rush of bulk orders. On some days, sales ran about five times higher than normal. Separately, trade records show that more than 50 tons of Japanese books were shipped to the United States since last year.
The story was first reported by Nippon TV’s news network, NNN, and later covered by Gadget Review and Tom’s Hardware. Here are the main facts from that reporting:
- The orders came from several different buyer accounts, spread across many stores.
- The books were all sent to one delivery center in Okayama Prefecture. The company that runs it did not answer reporters’ questions.
- Trade data from a company called Sayari showed that a company in the group of a major Japanese book distributor shipped more than 50 tons of boxes labelled “JAPANESE BOOKS” to the US.
- At about half a kilogram per book, that is roughly 100,000 books.
The type of books also stood out. A Tokyo bookseller told NNN that novels and comics used to be the big sellers. The new bulk orders were mostly philosophy, history, politics, medicine, law and books about life in Japan’s Edo period (the years 1603 to 1868).
What is not proven: no buyer has been publicly named. Nobody has shown for certain that the Okayama orders and the 50-ton shipment are linked, or that either one ended up with an AI company. The AI link is a strong suspicion, not a confirmed fact.
Why do AI companies want old books?
AI chatbots learn by reading huge amounts of text. Books are some of the best text there is: long, edited and checked for facts. Old books have one extra benefit. They were written before AI chatbots existed, so none of their words were made by a machine.
That last point matters more than it sounds. Since ChatGPT arrived in late 2022, the internet has filled up with AI-written articles, reviews and posts. When an AI learns mostly from other AIs’ writing, it can get worse over time. Researchers call this “model collapse.” A 2024 study in the journal Nature showed it happening: models trained again and again on AI-made text slowly lost the rare and unusual details and drifted into nonsense.
So a book printed in 1985 is a safe source. Every word in it was written by a person. As one book data company, ISBNdb, put it in its marketing to AI firms, the best AI training data “is sitting on a shelf.”
Old books also cover topics that barely exist online. A 1970s Japanese law textbook or a local history of an Edo-era town may never have been put on the internet. For an AI company, that is knowledge nobody else has.
Takeaway: old books are valuable to AI because they are human-written, carefully edited and often not online anywhere else.
This connects to a wider worry about where AI gets its information. We covered a related problem in how fake Reddit posts can change what ChatGPT tells you.
Why shred the books after scanning them?
The main reason is speed. The fastest way to scan a book is to cut off its spine and feed the loose pages through a high-speed scanner. Once the spine is gone, the book can’t be put back together. There is also a legal reason, explained below.
We know how this works because of a US court case. In Bartz v. Anthropic, authors sued Anthropic, the company that makes the Claude chatbot. Court papers made public in the case described an internal plan called “Project Panama.” According to those papers and reporting summarised on Wikipedia:
- Project Panama started in early 2024. An internal document called it an effort to “destructively scan all the books in the world.”
- Anthropic bought millions of used books from sellers such as Better World Books and World of Books.
- Machines sliced off the spines, the pages were scanned, and the paper was recycled.
- One vendor offered to handle up to 2 million books in six months, about 11,000 books a day.
- The planning document said the company did not “want it to be known” that it was doing this. The Washington Post reported on it in January 2026.
Anthropic is the only company whose book scanning has been confirmed in court papers. Other AI firms have not said whether they do the same.
Is it legal to scan and destroy books for AI training?
In the United States, one important court ruling says yes, as long as the company bought the book legally. Judge William Alsup ruled that buying a print book, scanning it, destroying the paper copy and using the digital copy to train AI counts as “fair use,” which is allowed under US copyright law.
The judge’s reasoning was simple. The company owned one paper copy. It turned that into one digital copy and threw the paper away. One copy replaced the other, and the digital version was not sold or shared.
Pirated books were a different story. Before Project Panama, Anthropic had also downloaded pirated copies of books from the internet. The judge did not accept that. Anthropic agreed to pay authors $1.5 billion to settle the case, about $3,000 for each of roughly 500,000 books.
That ruling helps explain the rush for paper books. Downloading books without permission can cost billions. Buying a used copy for a few dollars and scanning it is, at least under this US ruling, allowed.
Two limits are worth knowing:
- It is one ruling from one US court. Other courts, or other countries, could decide differently.
- Japan has its own rules. Japanese copyright law is known for being relatively open to using works for AI training, but these books were reportedly shipped to the US, so US rules would likely cover the scanning.
Should book lovers be worried?
Mostly no, but with one real concern.
Most books being bought were printed in large numbers. Cutting up one copy of a popular 1990s history book doesn’t make it disappear. Many other copies still exist in homes and libraries.
The real worry is rare and out-of-print books. If a buyer scans the last few copies of a small-print local history and then shreds them, the physical book is gone. The only surviving version may then be a private digital file inside a company that does not share it with the public. Rare-book dealers in Europe and the US have raised this same concern about unusual bulk orders for specialist titles.
There is also a question of fairness. Authors of old books usually get nothing when a used copy is resold, and nothing when that copy is used to train an AI. The US ruling does not change that.
If you want to know where this is heading, our piece on whether AI is building the next AI explains why AI companies are hungry for fresh, human-made data.
Frequently Asked Questions
Are AI companies really buying Japanese books?
Japanese used bookstores saw bulk orders at up to five times normal sales, and trade data shows more than 50 tons of Japanese books were shipped to the US. However, no buyer has been publicly named, so the link to AI training is suspected but not proven.
Why do AI companies want books printed before 2022?
Books printed before ChatGPT arrived in late 2022 were written entirely by people. Training AI on AI-written text can make it worse over time, a problem called model collapse. Old books are a safe, human-only source.
Is it legal to shred a book after scanning it for AI?
In the US, Judge William Alsup ruled in Bartz v. Anthropic that scanning a legally bought book and destroying the paper copy for AI training is fair use. Using pirated copies was not allowed, and Anthropic settled that part for $1.5 billion.
Which AI company was caught scanning and destroying books?
Court papers showed that Anthropic, maker of the Claude chatbot, ran “Project Panama” from early 2024. It bought millions of used books, cut off their spines, scanned the pages and recycled the paper. No other company has been confirmed in court to do the same.
Could rare books disappear because of AI scanning?
Common books are not at risk because many copies exist. The concern is rare or out-of-print titles. If the last copies are shredded after scanning, the only surviving version may be a private file the public can’t read.
