AI Firms Buy, Scan and Destroy Millions of Books

AI companies and intermediaries are buying large lots of used and out-of-print books, cutting or removing spines and pages, scanning high-resolution images for training datasets, and then discarding or pulping the remnants.

In recent months developers and data brokers have increased bulk purchases of secondhand inventory. Sellers who supply rare and obscure titles report that demand has risen sharply, with one bookseller reporting weekly sales rising from roughly 20 books to several hundred after buyers seeking training material began operating at scale.

Intermediaries marketing bulk sourcing services say they can locate foreign-language, overseas and hard-to-find titles that typically do not sell in standard used-book markets. Companies that acquire the physical copies often offer confidentiality to clients while delivering large numbers of digitized texts for model training.

Some industry participants argue that older, human-authored books are needed to provide clean examples of language and factual detail for large language models, particularly texts published before the widespread use of generative AI in 2023. Proponents say those works help maintain dataset quality as developers build larger models.

Courts have issued mixed rulings on the practice. A federal judge in San Francisco ruled last year in Bartz v. Anthropic PBC that scanning legally purchased books into digital copies, even when the originals were destroyed, can qualify as transformative fair use. Other federal judges issued similar fair-use findings in related cases involving major developers. In a separate decision in the same district, a settlement required Anthropic to pay about $1.5 billion to resolve claims by authors after the company used pirated digital copies in training, with payments estimated at roughly $3,000 per book for thousands of authors.

The buying and destructive scanning practice is reshaping the used-book market. Some sellers have earned significant short-term revenue from large sales, while collectors, librarians and some dealers say scarce and uncommon works are being lost to permanent destruction. A bookseller reported, ‘It benefits me financially as well as by clearing out old inventory that is otherwise unlikely to sell… I don’t like the end-use, and I don’t like that uncommon books are being pulped.’

Tech leaders and developers have debated how to handle physical collections. One executive wrote that teams should preserve rare volumes in libraries and scan them without cutting, rather than performing destructive scans.

With copyright litigation and public scrutiny ongoing, firms face legal and reputational risks as they decide how to acquire text for model training. Libraries and used-book dealers are monitoring changes in inventory and pricing as buyers pay premiums for obscure titles that collectors and researchers may value.