AI Companies May Be Destroying Rare Books to Train Their Models
Booksellers are raising the alarm that AI firms are buying up physical books, including rare and out-of-print copies, then shredding them to scan the pages cheaply and fast.
Key points
- AI companies are reportedly buying physical books, including rare titles, then destroying them to extract text for model training.
- Scanning bound books intact is slower and more expensive, making destruction the cheapest route to large volumes of high-quality text.
- Book dealers suspect the practice is more widespread than current reporting shows.
- Rare and out-of-print books destroyed this way could be lost permanently.
- Researchers and book lovers say the destruction is avoidable, if companies choose slower, less damaging methods.
Somewhere right now, a book that survived a century is probably being fed through a wood chipper.
That is the uncomfortable picture emerging from booksellers who, as first reported by Ars Technica AI, suspect AI companies are quietly buying up physical books and then destroying them to scan the pages quickly and cheaply.
Why would anyone destroy a book just to scan it?
Binding is the problem. A book's spine holds pages together in a curve, which makes flat, clean scanning almost impossible without pulling it apart first. Destroying the binding, cropping the loose pages, scanning them, and throwing the scraps away is dramatically faster than photographing an intact book page by page.
Speed matters here because AI companies are in a race. The models behind tools like ChatGPT and Claude, called large language models, learn from enormous amounts of text. Books are especially valuable for that training because they contain long, carefully written prose that is far richer than most text found online. Getting millions of pages into a training dataset quickly is worth money, and destroying a cheap paperback to do it is an easy call.
The harder call is what happens when the book is not cheap.
What about rare or irreplaceable books?
This is where the story turns genuinely troubling. Booksellers report that buyers, believed to be working for AI firms, are purchasing titles that include rare and out-of-print copies. Once those books are shredded for scanning, no physical copy may remain.
A digitised scan is not the same thing as a surviving artefact. Pages carry marginalia, printing variations, and physical evidence that scholars use. Losing the object means losing more than the words.
Dealers also suspect the scale is larger than what has surfaced publicly. Because bulk book purchases are not regulated or tracked in a way that reveals the buyer's purpose, the full extent is difficult to confirm.
Should ordinary readers care?
Yes, even if you never visit a rare-book shop. The books being scanned shape what AI tools know and how they write. If the most valuable, carefully crafted long-form writing goes into these models without the authors' or publishers' consent, and the physical copies vanish in the process, both the cultural record and the livelihoods of writers take a hit.
The cruelest part, as critics point out, is that none of this is technically necessary. Slower, non-destructive scanning methods exist. Companies are choosing the cheaper path.
What to watch for: If you sell second-hand or rare books, be alert to bulk buyers who ask few questions about condition and show no interest in resale. Unusually large orders for reading copies of out-of-print titles, paid quickly and collected in bulk, fit the pattern dealers are describing.
