AI companies are scanning and destroying old books

AI companies are buying old books, scanning them and destroying the remains, writes Emanuel Maiberg at 404 Media. The program is enabled by ISBNdb, maintainers of a vast book database, which helps them buy vintage physical media in bulk: "The world's best AI training data is sitting on a shelf," its pitch reads.

Unlike text found online and in recently-published books, the old stuff is guaranteed to be free of slop: a hazard in AI training likened to inbreeding. ISBNdb can arrange buys of up to a million books, it claims, from libraries, retailers and estate sales. ISBNdb promises confidentiality, too, warning clients that "the optics problem is real."

"'AI company destroys two million books' is not a headline that generates sympathy," it helpfully relates.

Bulk book sellers confirm surgees in large orders in the last year. One told 404 Media the buys clear dead inventory (unsold stock unlikely to sell for a profit) but also that "uncommon books are being pulped."

Destructive scanning cuts off the spines of books, feeds pages through the machine, then bins the remains. It's faster and cheaper than non-destructive scanning, as practiced by real libraries and the Internet Archive.

— Jason Scott (@textfiles.com) 2026-07-29T18:58:29.266Z

Training large language models on books is not quite settled law, but Bartz v. Anthropic, in which Judge William Alsup ruled that it represented a transformative fair use, was the green light that AI companies hoped for. But Alsup determined that Anthropic must pay $1.5 billion for pirating and storing digital copies of the books, which led to the explosive interest in bulk physical purchases and destructive scanning.

AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop [404 Media]