Rare books are being destroyed for AI training

AI companies are reportedly buying and destroying rare printed books to build training datasets, with ISBNdb shipping up to one million books anonymously to AI laboratories. The practice involves removing book spines, separating pages, and feeding them through high-speed scanners that can make digitization faster and cheaper—but leave the original volume unusable.

Why pre-2022 books are sought for training

Reports from 404 Media found that ISBNdb supplies physical books to AI developers looking for human-written material unaffected by modern chatbot output. The company argues that books published before 2022 provide cleaner data because they predate the widespread use of AI-generated text.

These works are described as “dense, edited, authoritative,” in contrast with online material that increasingly contains machine-generated content of uncertain quality. Using older books may also help companies avoid model collapse, a decline in quality that can occur when models repeatedly train on synthetic text produced by earlier models.

ISBNdb also says older printed works are less likely to contain deliberate data-poisoning techniques, in which authors modify documents to interfere with AI training. The company reportedly supplies books under strict non-disclosure agreements, keeping its commercial customers' identities confidential.

Rare books and the preservation problem

Booksellers interviewed by 404 Media said some volumes entering scanning programs have only a handful of surviving copies after wars, fires, and centuries of handling. Unlike widely available websites or modern publications, these historical works may be impossible to reproduce once their physical copies are destroyed.

Recommended reading

Amazon completes $50 billion OpenAI deal early

ISBNdb reportedly acknowledges the reputational risk of the process. Its website states:

“AI company destroys two million books” is not a headline that generates sympathy.

A recent United States court ruling involving Anthropic found that scanning legally purchased books for AI training constituted fair use under specific circumstances. Part of the reasoning was that destroying each printed copy during scanning meant one legal copy effectively replaced another, rather than creating multiple copies.

In response to a critic identified as @Hedgie on X, Elon Musk said:

“I’ve asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way.”

The source does not identify which AI companies are receiving the books or how many rare works have already been destroyed. It reports that ISBNdb’s method can involve anonymous bulk shipments and that the company’s own example refers to “two million books,” but it does not establish that two million books have actually been destroyed.