Amazon Destroys Rare Books to Train Nova AI Models
An investigative tracking effort revealed that Amazon is purchasing and destroying physical books to digitize their contents for training its Nova artificial intelligence models.

An investigation by 404 Media has exposed how Amazon acquires physical books, digitizes them, and destroys them to gather training data. By placing an AirTag inside a shipment of rare books, investigators tracked the package to an Amazon facility in Las Vegas. This warehouse is home to an internal team known as VGT3, whose logo depicts a Tyrannosaurus rex clutching a book. Workers at the facility reportedly slice off book spines to accelerate the scanning process, rendering the physical copies useless.
Amazon uses the resulting digital text to train its proprietary Nova AI models. A company spokesperson confirmed that Amazon acquires books through standard commercial channels to refine its products. Industry experts and booksellers believe AI developers are attempting to systematically digitize every book by its ISBN number. Physical books are highly prized for AI training because they frequently lack digital equivalents and predate 2022, ensuring the text is entirely free of synthetic, AI-generated content.
Amazon is not the only technology giant employing this destructive digitization strategy. A copyright lawsuit previously exposed Anthropic's "Project Panama," a similar initiative where the startup purchased physical books from online marketplaces, severed their spines, and scanned them. In that legal battle, a judge determined the practice constituted fair use under copyright law, reasoning that destroying the physical originals prevented them from being illegally duplicated and resold.
While legally defensible in some jurisdictions, the practice has sparked intense criticism from preservationists and authors. Critics argue that destroying rare and sometimes irreplaceable physical volumes permanently removes knowledge from the public sphere. Instead of remaining accessible on library shelves, this information is effectively locked away inside the proprietary, closed-source models of a single technology corporation.
This is our own summary of reporting by The Decoder



