Amazon is destroying rare books to train its AI models, investigation finds

Amazon Amazon.com, Inc. ↘ 1.36% is buying large quantities of rare and out-of-print books, slicing off their spines for faster scanning, and destroying the originals — all to feed training data into its AI models. An investigation published Monday by 404 Media revealed the practice after the outlet placed an Apple AirTag inside a shipment of roughly 1,000 books and tracked it to an Amazon warehouse in Las Vegas.404media

"Amazon purchases books through commercial channels to help develop and improve the products and services our customers use," an Amazon spokesperson said in a statement, declining to answer questions about how many books it has destroyed or how many such facilities it operates.404media

Inside VGT3

The tracked book traveled from California through Wisconsin and Colorado before arriving at an Amazon facility called LAS8 in Las Vegas. Workers there describe a separate operation housed at the site's north end, coded VGT3, whose logo is a Tyrannosaurus rex clutching a book.404media

Employees say their sole job is receiving massive shipments of printed books, cutting off their bindings, and scanning the pages. "I work at VGT3 here in Vegas, and all we do is scan books," one worker wrote on an Amazon employee forum. "Some are assigned to cut books, and others go to receive where they get books and scan the bar codes."404media

Booksellers told 404 Media they suspect AI companies are working through the global catalog of ISBNs to scan every published book systematically. The bulk orders they receive are not price-sensitive and include seemingly random titles, unlike purchases by libraries or collectors.404media

Why Physical Books

Pre-2022 printed books are prized because they are guaranteed free of AI-generated text. When language models train on synthetic output, they risk a degradation known as "model collapse." Many of these texts also contain information not available anywhere on the internet, making them an untapped reservoir of training data.techcrunch+1

The practice follows the same template revealed in Anthropic's "Project Panama," exposed through litigation by book authors. In that case, a federal judge ruled in June 2025 that scanning lawfully purchased books for AI training constitutes fair use — in part because the buyer destroyed the original copy.gadgetreview+1

A Broader Controversy

The revelation arrives as Amazon restructures its AI strategy around frontier-model research. The company has wound down several in-house Nova models and consolidated efforts under a Frontier Model Research group led by Pieter Abbeel, following the departure of former AI chief Rohit Prasad in December 2025.firstpost

For booksellers, the calculus is simpler. "There are different types of value," the anonymous bookseller who helped 404 Media said. "There's historical value, intellectual value, sentimental value. All sorts of things, and all of those the AI companies don't care about. They just want the content as a bunch of words strung together."404media