Hidden AirTag exposes Amazon's destruction of rare books for AI training
An AirTag secretly planted inside a rare book led 404 Media to an Amazon warehouse in Las Vegas where workers scan and then destroy bulk-ordered rare books to build AI training data. Amazon has declined to directly confirm the AI training purpose.
For roughly a year, booksellers have suspected that AI companies were buying up large lots of rare books and destroying them after scanning, using the text to train AI models. That suspicion has now been confirmed by 404 Media, which worked with a bookseller to hide an AirTag inside a rare book that was part of a bulk order. The tracker led investigators to an Amazon AI training facility in Las Vegas, where a team pulls books apart at the spine and scans their pages.
The warehouse door for this internal team, known as VGT3, reportedly displays a logo of a Tyrannosaurus rex about to devour a book.
Amazon avoids specifics
Amazon declined to comment directly on the findings, instead repeating a general statement that it purchases books through commercial channels to help develop and improve products and services for customers. The statement made no mention of AI training. However, Amazon is known to be building what it considers frontier AI models, which require enormous volumes of unique training text to compete with companies like Google, OpenAI, and Anthropic — rivals that have reportedly said publicly they do not train on rare or antique books.
Online forum discussions among Amazon workers, reviewed by 404 Media, suggested the company ran short of books to scan earlier this year, raising fears the facility might close. At one point the supply reportedly ran out entirely. The warehouse remains operational, and bulk orders of rare books delivered there are now confirmed to be routinely destroyed.
ISBN targeting and ethical concerns
The investigation also lends support to another bookseller theory: that AI firms are systematically hunting for books by their ISBN numbers to maximize the number of unique works represented in their training data. Workers reportedly were trained to check barcodes or ISBNs before scanning.
The books targeted are typically not prized first editions but lower-value, obscure works — such as texts never translated from a lesser-used foreign language or books that were never widely distributed. The bookseller who planted the AirTag noted that such works can still hold historical, intellectual, or sentimental value that AI companies overlook, since they are primarily interested in text as raw content.
Online commentary on the practice has been mixed. Some argue that many of these books would otherwise sit unsold on shelves, while others warn that once scanned, the content effectively disappears from public access, since AI firms are unlikely to share what they've digitized. A Scottish bookseller quoted separately by the BBC argued that AI companies should distinguish between texts that are truly replaceable and rare surviving editions, calling destruction of the latter a far more serious problem.
