AI Companies Are Buying Used Books in Bulk to Train Their Models

A new report from 404 Media says AI companies are buying used books in large amounts to get good training data. They want books printed before 2022, because those are less likely to be full of AI-generated junk. Orders range from 1,000 to 1 million books. Used book sellers say sales have jumped up to five times more than usual.

These companies use a method called destructive scanning, which tears the books apart for fast digitization. This has already destroyed millions of books. Earlier, Anthropic had to pay $1.5 billion for using a database of pirated books to train its Claude model. Google is also being sued by publishers for similar issues.

Critics worry that rare books could be destroyed in this process. Also, the scanned content goes into private AI databases, so the public can't access it anymore. This raises a big question: will we get smarter AI at the cost of losing access to human knowledge?