AI Companies Are Shredding Rare Books to Feed Their Models

Companies like Anthropic and OpenAI are bulk-buying printed books and destroying them in the scanning process, alarming rare book dealers who say irreplaceable editions are vanishing into the training-data pipeline.

AI companies desperate for fresh, human-written training data have found a new hunting ground: the world's libraries and used-book warehouses. And they're not borrowing — they're shredding.

According to a 404 Media investigation, ISBNdb — a company that bills itself as "the world's largest book database" — is now offering high-volume book acquisition services explicitly aimed at AI firms. The pitch? Printed books are guaranteed free of AI-generated slop: "curated, peer-reviewed, domain-specific human knowledge, structured in a way no web crawl can replicate."

The economics are brutal. Destructive scanning — where a book's spine is sliced off and pages are fed through high-speed scanners — is significantly faster and cheaper than non-destructive digitization. One procurement list obtained by journalists contained 3,000 English-language titles, from academic works like Laser Shock Peening of Advanced Ceramics to irreplaceable folklore studies. Once scanned, the physical books are mulched and recycled.

Rare book dealers in the Netherlands raised the alarm in June, warning that obscure academic editions — some with print runs under a thousand copies — are being bought in bulk with no mechanism to check whether a digital copy already exists. The practice is entirely legal. Under US copyright law, companies can purchase physical books, convert them to digital text for training, and destroy the originals. But as commenters on Hacker News pointed out, the parallels to Vernor Vinge's dystopian "shred and scan" factories from his novel Rainbows End are no longer fiction.

The underlying tension is familiar: copyright law was designed for an era when copying meant a physical duplicate. It never anticipated corporations buying the world's books to feed them into neural networks, then pulping the evidence. Whether that qualifies as "fair use" is a question courts are only beginning to answer — while the shredders keep running.

Sources: 404 Media, HedgieMarkets via xcancel