Amazon Destroys Rare Books After Scanning Them for AI Training

Amazon is purchasing large quantities of books, scanning them to train artificial intelligence models, and destroying the physical copies in the process, according to an investigation by 404 Media that tracked a shipment of rare books to an Amazon facility in Las Vegas.

The operation, which has not been previously reported, takes place at a warehouse called VGT3 where employees receive mass shipments of printed books, cut off the bindings to speed scanning, and discard the originals. The facility's logo features a Tyrannosaurus rex holding a book in its jaws.

"Amazon purchases books through commercial channels to help develop and improve the products and services our customers use," an Amazon spokesperson told 404 Media.

How the investigation worked

404 Media placed an Apple AirTag inside a shipment of approximately 1,000 books sold through Biblio, an online marketplace that keeps buyer identities anonymous. The tracked book traveled from California to Milwaukee, then to a distribution center near Kenosha, Wisconsin, before heading west by truck through Colorado and finally arriving at Amazon's LAS8 warehouse in Las Vegas.

The north end of LAS8 houses the VGT3 operation. Amazon employees who work there describe their jobs as repetitive but straightforward: scanning book barcodes (ISBNs) and then scanning the content itself.

"I work at VGT3 here in Vegas, and all we do is scan books," one Amazon employee wrote on a worker forum. "Some are assigned to cut books, and others go to receive where they get books and scan the bar codes."

The practice of scanning ISBNs—unique identifiers assigned to every published book—supports booksellers' theory that AI companies are systematically working through lists of ISBNs to scan every printed book in existence.

Why it matters

Printed books represent valuable training data for AI companies because much of their content isn't available online and is guaranteed free of AI-generated text if published before 2022. Training models on AI-generated content can cause "model collapse," degrading performance. Amazon is developing its Nova family of large language models to compete with OpenAI, Google, and Anthropic, all of which require massive text datasets. The destruction of physical books, even if legal under fair use doctrine, raises questions about preserving cultural artifacts for purposes beyond their immediate commercial value. A judge previously ruled that Anthropic's similar book-scanning operation constituted fair use partly because the company destroyed originals, preventing duplication.

Booksellers first noticed unusual bulk purchases starting last year—large orders of seemingly random titles from buyers who showed no price sensitivity, unlike typical institutional purchasers such as libraries. The books tracked in this investigation are rare, though not necessarily valuable in monetary terms, representing works in foreign languages or limited print runs that few people seek but that hold historical and intellectual value.

Amazon did not respond to questions about how many such facilities it operates, why it destroys books rather than using non-destructive scanning methods, or its response to public criticism of the practice.

The details were first reported by Emanuel Maiberg at 404 Media.