Pulp nonfiction

TL;DR: For months, booksellers wondered why they were receiving bulk orders for a bunch of obscure used books. Now they may have an answer: An Amazon facility in Las Vegas appears to be processing used books and digitizing them, presumably to train its AI.

What happened: 404 Media says it got a bookseller to plant an Apple AirTag in a book that was a part of a large order. The AirTag took a rather circuitous journey, from California to Milwaukee, then Colorado, before it arrived at an Amazon facility in Vegas.

That facility contains a previously unknown warehouse called VGT3, marked by a logo of a menacing dinosaur gripping a book. There, workers spend their time scanning rare books—sometimes destroying them in the process.

We don’t know much else about the operation: An Amazon spokesperson only confirmed to 404 that the company “purchases books through commercial channels to help develop and improve the products and services our customers use.”

The “rare” used books mystery: Booksellers noticed the unprecedented uptick in used book sales because of the order sizes and the kind of books being purchased. The texts are considered “rare” because they cover niche topics and few copies remain—an example being How to Use Corel WordPerfect 1991, per one bookseller. So we’re not talking about a first-edition James Joyce. Still, the trend couldn’t definitively be connected to AI firms until now.

Not just Amazon: We know that other AI firms have run a similar playbook—Anthropic drew scrutiny earlier this year for buying and ripping out the spines of millions of print books for AI training. Both Amazon and Anthropic presumably cut up books because it makes scanning pages faster than turning them individually, and buying and destroying the books may provide some legal cover for AI companies under fair-use laws.

Why this matters: AI labs are snapping up print books as untapped sources of training data dry up. The internet has been scraped already—and it’s quickly filling up with AI-generated text, which can in turn degrade the quality of AI models.

These rare books provide fresh material, but once they’re scanned and destroyed, AI companies are the only ones with access to the content within. Reactions online have been predictably heated (Elon Musk assured users on X that he asked his team to scan books “the hard way”). Workers at VGT3, meanwhile, describe the job as one of the more desirable, less stressful warehouse roles at Amazon.

Bottom line: This used books operation reveals the ingredients that go into the AI sausage: a whole lot of human-written text that gets blended into one big dataset, with the authors’ names filed off. —WK