Amazon facility reportedly destroys books by spine removal to scan for AI training data

An investigation by 404 Media used an Apple AirTag hidden in a rare book to trace its journey to an Amazon facility in Las Vegas, identified as VGT3. The book was part of a 1,000-book shipment purchased through Biblio, an online marketplace for independent sellers. The facility’s sole reported function is to remove book spines and scan the pages, a process that destroys the physical copies. This suggests a systematic approach to converting books into digital data, likely for AI training purposes.

The method of spine removal and scanning contrasts with earlier book digitization efforts, such as Google Books, which used non-destructive scanners to preserve physical copies. The shift to destructive scanning may reflect cost or efficiency pressures, as industrial scanners can process flat pages faster than bound books. However, this approach sacrifices the integrity of the original material, raising questions about the long-term availability of rare or out-of-print books used in these operations.

Employees at VGT3 reportedly scan the ISBN of each book, supporting the theory that AI companies are systematically working through lists of published titles. Large orders from marketplaces like Biblio often include a mix of genres and conditions, suggesting a focus on volume rather than curation. The lack of rare books without ISBNs in these orders further indicates a targeted, data-driven acquisition strategy.

The investigation aligns with broader reports of AI companies acquiring books in bulk, sometimes through questionable means. Earlier legal cases, such as Anthropic’s $1.5 billion fine for pirated books, highlight the legal and ethical risks of these practices. For engineers, this underscores the tension between data acquisition and preservation, as well as the need to consider the provenance and handling of training datasets in AI development.

The existence of a dedicated facility like VGT3 signals a shift in how AI training data is sourced and processed. While the scale of operations may accelerate model development, it also introduces logistical and ethical challenges. Engineers working on AI systems should be aware of these practices, as they may influence data quality, legal compliance, and public perception of AI projects.