An AirTag hidden inside a shipment of rare books just exposed something ugly about how AI companies source their training data. 404 Media tracked the package across the country until it landed at an Amazon warehouse in Las Vegas, where a team called VGT3 — their logo is a T. rex clutching a book, which now reads as darkly ironic — processes printed texts for AI training. Workers there reportedly slice off book spines to make scanning faster, a process that destroys the physical copies entirely.
The scanned content feeds Amazon’s Nova models. A spokesperson confirmed the company buys books through commercial channels to improve its products, which is technically true and strategically vague. What the statement leaves out is that the books being destroyed are sometimes rare editions that may never be replaced. Booksellers tracking the pattern believe AI companies are methodically working through ISBN databases, hunting for printed texts that don’t exist online and predate 2022 — meaning they’re free of AI-generated content and therefore uniquely valuable as clean training data.
Amazon isn’t the only player here. Anthropic ran a nearly identical operation called Project Panama, exposed through a lawsuit by book authors. The company bought books on marketplaces, cut off their spines, and digitized them. A judge ruled the practice qualified as fair use, partly because destroying the originals meant they weren’t being copied and resold. That logic — that destruction makes infringement more legally palatable — is the kind of reasoning that makes copyright lawyers wince and book lovers furious.
The broader pattern is troubling. Both companies are converting physical knowledge that belonged to the public commons into private training material locked inside closed models. Unlike the early days of AI training, when companies scraped the open web and at least the source material remained accessible, this approach permanently removes the originals from circulation. A used bookstore in Oklahoma can’t resell a book that’s been spine-cut and fed into a scanner in Nevada. The information survives, but only as weights inside a model that no one outside the company can ever inspect.
