Amazon is acquiring large quantities of rare and antique physical books through commercial channels to scan and use the text for training its artificial intelligence models [1, 2, 3]. These books, which include rare, out-of-print, and pre-2022 printed editions, are especially valued because their content is not available online and uncontaminated by AI-generated text, avoiding degradation in model performance known as "model collapse" [1, 2, 3].
During the scanning process, Amazon physically destroys the original books. This involves slicing off their bindings to speed the scanning and then shredding the volumes [1, 2, 3]. Employees at an Amazon warehouse facility labeled VGT3, located in Las Vegas, report that their primary job is scanning these books for AI training purposes [1, 3].
A recent investigation by 404 Media tracked a shipment of about 1,000 rare books arriving at the VGT3 warehouse on August 17, confirming the scale of the operation [1, 3]. This surge in physical book sales first appeared in July 2026 with reports suggesting bulk purchases by AI companies to obtain high-quality training data [1, 3].
The motivation for buying physical copies instead of digital versions is partly to avoid legal risks related to unauthorized digital copies and to gain access to text not available online [3]. Other AI firms, including Anthropic, have also been revealed to purchase millions of physical books for similar purposes, as shown by court documents about a program called "Project Panama" [3].
An Amazon spokesperson said, "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use" [1, 2].
The destruction of these rare physical books and their conversion into AI training data highlights how the industry sources unique and clean datasets. The investigation continues to track shipments and operations at Amazon's VGT3 facility.
