← Back to digest
Industry

Amazon is destroying rare books to train AI

What's this about?

404 Media planted a tracking device inside a rare book it suspected would be bought by an AI company for training data, then followed it across the country. It ended at an Amazon warehouse in Las Vegas. Employees there describe the work: receive enormous shipments of printed books, cut the bindings off so the pages feed through a scanner faster, and destroy the physical book in the process. The team's logo is a dinosaur. The publication deliberately withheld the titles, and noted the shipment was of genuinely rare books — few copies in circulation, sometimes in languages with little digitized text.

What exactly is "destructive scanning"?

Scanning a bound book is slow and awkward: you open it, flatten the spine, photograph two curved pages, repeat. Cut the spine off and you have a stack of loose sheets you can run through a document feeder like any other paper — dramatically faster and higher quality. The book does not survive. Libraries have used this for decades for damaged material, but always as a last resort for items with surviving copies. Applying it at industrial scale to rare books inverts that logic: the scarcer the book, the more valuable its text as training data, and the more consequential its destruction.

What's the community saying?

HN 133 points, 291 comments — plus Ars Technica and TechCrunch pickups
  1. The rarity is the point, and commenters worked out why. These models have already ingested essentially the whole public web. What's left of value is precisely the text that was never digitized — obscure books and undigitized languages — which is also the text with no backup.
  2. "Bad enough to steal IP, but destroying books — that's just evil." The emotional core of the thread, and the reason this story travelled beyond the usual audience.
  3. A genuine mitigating context. Several noted the scanning facility predates the AI push; as the world's largest bookseller, Amazon has long had reasons to hold digital copies of everything, so this may be an existing operation repurposed rather than built for training.
  4. Skepticism about the marginal value. A recurring technical question: with the entire internet already trained on, what does an obscure book actually add? The answers pointed at exactly the gap the article identifies — low-resource languages and material no scrape ever reached.
  5. Criticism of the messenger. A well-upvoted complaint about 404 Media's style as built for outrage — a reminder that this thread was not uniformly on the publication's side.
← Red Agent meets Copilot Autofix at SnowflakeAI;DR (AI; Didn't Read) →