Updated
Updated · PC Gamer · Jul 28
AI Companies Buy Up to 1 Million Books Anonymously, Destroying Them for Training Data
Updated
Updated · PC Gamer · Jul 28

AI Companies Buy Up to 1 Million Books Anonymously, Destroying Them for Training Data

3 articles · Updated · PC Gamer · Jul 28

Summary

  • 404 Media reported AI companies are using middlemen to anonymously source physical books at industrial scale, then scan and discard them for model training.
  • ISBNdb had advertised book-sourcing services tailored to LLM training, saying clients could buy up to 1 million books per order under strict NDAs; the company later removed the page and said it was pivoting away.
  • The tactic appears aimed at avoiding backlash while securing pre-LLM, human-written text that is cleaner and rarer than internet data increasingly flooded with AI-generated material.
  • Anthropic was already shown in a 2025 court ruling to have destroyed millions of print books while training Claude, and internal documents said the company did not want that work publicly known.
  • Booksellers say demand has surged since April, with one rare-book dealer now filling hundreds of orders a week and seeing uncommon titles pulped after purchase.

Insights

If AI companies destroy physical books to hide their data sources, how will regulators ever audit what these models actually know?
What happens when the AI industry's secret appetite for clean data leads to the quiet destruction of our rarest physical books?
Could the very books being shredded to build tomorrow's AI be the last remaining records of untainted human history?