Key Points

  • An Apple AirTag concealed within a significant book order led to an inquiry that traced the shipment to Amazon's VGT3 facility in Las Vegas.
  • At this location, employees dismantle book spines to facilitate quicker scanning, resulting in the destruction of the physical books.
  • Amazon acknowledged that it "purchases books through commercial channels" and noted that this operation had not been previously disclosed.

AI firms are processing large quantities of books, including some that are rare, often resulting in their destruction during the training of AI models. Recent findings shed light on the companies involved and their methods.

A report from independent tech outlet 404 Media utilized a tracking device placed in a shipment of rare books to monitor its journey to an Amazon facility in Las Vegas, where the company scans and ultimately destroys printed texts for AI training purposes.

Myriad: When will OpenAI release GPT-6? Click to make your prediction.

This investigation, published on Monday, is the first to publicly identify where bulk book acquisitions related to AI training ultimately conclude.

The final destination was the VGT3 team at Amazon, where employees revealed that their primary task is to receive large deliveries of printed books and then cut the bindings to allow for quicker scanning. This process results in the destruction of the books. Notably, the team's logo features a dinosaur grasping a book.

Image: 404 Media

The Rising Demand for Printed Books

Books published before 2022 contain text that is not widely available online and is crucially not generated by machines. Training AI on its own outputs can lead to "model collapse," where quality deteriorates with each iteration.

Thus, pre-2022 printed material serves as an ideal resource for AI training.

Amazon is not alone in this practice. A burgeoning industry has emerged to provide AI laboratories with physical books that are stripped, scanned, and discarded. This process mirrors the dystopian imagery of Ray Bradbury's "Fahrenheit 451," where texts are destroyed to feed machines. Anthropic's internal "Project Panama" already operates at scale, digitizing millions of books through this destructive scanning approach.

This quest for clean text has led to legal battles. A federal judge ruled that training AI on legally acquired books can be considered fair use, marking a partial victory for Anthropic, a stance echoed by Meta and OpenAI. However, issues surrounding pirated copies are treated differently; Anthropic recently settled for $1.5 billion over scanned pirated works, and Salesforce is now facing a class action lawsuit regarding alleged book piracy.