AI Companies are buying old books, cutting them apart, scanned by AI models and then throwing them away. Here is what 404 Media found, what ...
Here is what 404 Media found, what has been confirmed, what has been denied, and why old books suddenly matter so much to AI labs. I saw a screenshot going around saying AI labs are destroying rare books to train their models. It sounded like one of those posts that falls apart when you check it. So I checked. Some parts held up, some parts were walked back, and the full story is more interesting than the screenshot.
AI Companies are buying old books, cutting them apart, scanned by AI models and then throwing them away.
Where This Started
The story comes from 404 Media, an independent tech news outlet. Their reporting centered on a company called ISBNdb, which calls itself the world's largest book database. For years it was mostly known as a place where libraries, bookstores, and apps get information about books. According to 404 Media, ISBNdb had started offering to source physical books in bulk for AI labs to scan into training data. Orders reportedly ranged from a thousand copies all the way up to a million books in one go. The pitch on its website was simple. It said the best AI training data is already sitting on a shelf.
The method is not complicated. A book gets bought, its spine gets cut off, the loose pages go through a fast scanner, and what is left gets thrown out. Cutting the spine is the key step. A bound book has to be opened page by page, which is slow. Loose pages can be fed through a machine very quickly. So the book is sacrificed for speed.
Why Old Books
This is the part I found most interesting, because it shows a real problem inside AI. The internet is now full of text written by AI. If a new model trains on that text, it is partly learning from older models instead of from people. Researchers call the long-term result "model collapse," where quality slowly drops because models keep feeding on their own output. Old books avoid that problem. ISBNdb argued that anything published before 2022 is ideal because it cannot contain AI-written text. 2022 is roughly when ChatGPT arrived, so anything printed before then is guaranteed to be written by people. There is also a quality argument. The Washington Post reported earlier this year that an Anthropic co-founder had suggested books could help models learn to write well, instead of copying the low-quality style common online. Books are edited, structured, and usually better written than random web pages. For a company trying to build a better model, that is valuable.
One detail made this story spread. ISBNdb apparently understood how it would look. Its site openly admitted there was an optics problem, and noted that a headline about an AI company destroying two million books would not win anyone sympathy. According to 404 Media, every project started with a non-disclosure agreement. So the buyers could stay anonymous, and you could not easily tell which lab was behind a big order.
This was not only a claim on a website. People who sell books saw it happening. One bookseller told 404 Media they used to sell around 20 books in a good week, and in recent months that jumped to several hundred. Sellers on Alibris and Biblio reported similar spikes. It was not only in the US either. Booksellers in the Netherlands, Germany, Switzerland, and Spain reported large orders for niche and academic titles. One Dutch seller got a request for nearly 3,000 books from a company based in Singapore. Most of what was being bought seems to be unsold stock. But not all of it. One seller said that while the orders cleared out old inventory, some uncommon books were also being pulped.
Then ISBNdb Backed Away
About a week after the story, ISBNdb changed its position. It took down its AI pages, said it had never bought, scanned, or sold a single book for AI training, and described the whole offer as a test of market interest. Its statement said the service was never actually launched and that its normal database business was unaffected. So to be fair, the claim that ISBNdb shipped millions of books to AI labs is now disputed by ISBNdb itself. The screenshot that went viral does not mention this.
The Tracking Device
This is where the story got much more solid, and it did not depend on ISBNdb at all. 404 Media wanted to know where these bulk orders actually end up. Working with a bookseller, they hid an Apple AirTag inside an order of about 1,000 books sold through Biblio, then followed it. The route was long. It went from a California airport to Milwaukee, then to a warehouse in Kenosha, Wisconsin, then through Colorado Springs and a truck stop in Grand Junction, and finally to an Amazon site in Las Vegas called LAS8. Inside that site is a team called VGT3.
Workers there said, in posts to each other online, that their whole job is cutting the bindings off books and scanning them, sometimes for up to 12 hours at a time. One detail almost sounds made up, but it was verified. The team's logo, visible at the facility entrance, is a dinosaur holding a book in its claws. Amazon did not deny buying books. Its response was a single sentence saying it buys books through normal commercial channels to improve its products and services. Amazon builds its own large AI models called Nova, which it describes as frontier level.
Why This Is Legal
A lot of people assume this must be illegal. It currently is not, and the reason is a court case involving Anthropic. In June 2025, in a case called Bartz v. Anthropic, Judge William Alsup ruled that scanning books that were legally bought and using the scans to train AI counts as fair use. Part of his reasoning was that the digital copy replaced the physical one, rather than creating extra copies or sharing them around. That lawsuit is also how people first learned about the scale of this. The judge ordered Anthropic to unseal internal documents about a book scanning effort called Project Panama. That was the first public sign that AI companies wanted to scan millions of physical books.
Former US congressman Brad Carson made a point about this ruling that I keep thinking about. His argument was that the court approved the scanning partly because the original book gets destroyed, so one legal copy simply replaces another. If that reading is right, the law ends up favoring destruction over preservation. A company that scans a book and throws it away is on safer legal ground than one that scans it and keeps it or gives it to a library. Nobody designed it that way on purpose, but that is the effect.
What Is and What Is Not
It is worth separating the facts, because the viral version mixes them up. Confirmed: AI companies want old printed books because they are free of AI-written text. Bulk orders to booksellers went up sharply. A tracked shipment ended at an Amazon facility where workers cut and scan books. Scanning purchased books for AI training has been ruled fair use in the US. Disputed: ISBNdb now says it never actually bought or sold a single book and that its offer was only a market test. Not proven: that rare books are being specifically targeted. Snopes left this part unrated, saying it is still unclear whether Amazon goes looking for rare books or whether the books are used only for AI training. Most of the destroyed books are probably common, unsold copies. Some uncommon ones clearly got caught up in it too.
Two Ways to See It
I do not think this has one obvious answer. One view is that nothing important is lost. Most of these books were unsold stock that might have been pulped anyway, and their content now lives on digitally. Some people argue that if you had to choose between saving the paper and letting the knowledge be used at scale, the paper is the smaller loss. The other view is that physical books are more than their text. An old edition can carry notes, printing details, and history that a scan does not keep. When a title has only a few copies left, shredding one is not a small thing. And the scans are not going to a public library. They go into a private company's training data, which the public cannot access. What bothers me most is not the scanning itself. It is that the people doing it knew how it would look, used anonymous buying and NDAs to avoid being noticed, and only got caught because a journalist put an AirTag inside a box of books.
Resources
404 Media, "Company Offering Printed Books to Train AI Stops After 404 Media Coverage": https://www.404media.co/ai-company-training-scanning-books-database-isbndb/
404 Media, "We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility": https://www.404media.co/we-tracked-a-shipment-of-rare-books-it-ended-at-an-amazon-ai-training-facility/
The Next Web, "AI firms are buying old books for slop-free training data": https://thenextweb.com/news/ai-companies-buying-old-books-training-data-slop
Tom's Hardware, "AI companies are reportedly shredding millions of books after using them to train AI models": https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-companies-are-reportedly-shredding-millions-of-books-to-train-models-tech-giants-outsource-to-middlemen-to-secretly-buy-up-books-for-training-material
Tom's Hardware, "Secret tracking device placed in rare book ends up in Amazon processing facility": https://www.tomshardware.com/tech-industry/artificial-intelligence/secret-tracking-device-placed-in-rare-book-ends-up-in-amazon-processing-facility-destroying-books-to-train-ai-models-is-all-the-vegas-warehouse-does
Snopes, "Are AI companies scanning and destroying millions of books, including rare titles?": https://www.snopes.com/fact-check/ai-companies-destroying-rare-books/
Snopes, "Is Amazon buying and destroying large amounts of books to train AI?": https://www.snopes.com/news/2026/09/01/amazon-destroying-books-ai/
Jerusalem Post, "AI burns books? Tech giants are buying, shredding books to train smarter chatbots": https://www.jpost.com/business-and-innovation/tech-and-start-ups/article-904336
Boing Boing, "AI companies are scanning and destroying old books": https://boingboing.net/2026/07/30/ai-companies-are-scanning-and-destroying-old-books.html
Dallas Express, "The Vanishing Page: AI Firms Scan Then Destroy Rare Book Editions": https://dallasexpress.com/national/the-vanishing-page-ai-firms-scan-then-destroy-rare-book-editions/
AI Weekly, "Amazon Scans and Destroys Rare Books at Vegas AI Facility": https://aiweekly.co/alerts/amazon-scans-and-destroys-rare-books-at-vegas-ai-facility
