Asger Risom
Tech companies are turning to “destructive scanning”—secretly buying thousands of physical books, shearing off their spines, and digitizing the pages—to bypass the limitations of internet scraping and train next-generation AI models on high-quality academic texts.
When Dutch antiquarian bookseller Pieter de Vries received an email requesting over 3,000 academic titles to be shipped to China, he assumed it was a standard phishing scam.
The reality was far stranger: it was a rare glimpse into the physical supply chain of the artificial intelligence industry, where tech companies are secretly mass-purchasing printed books, cutting off their spines, and destroying them to train next-generation AI models.
The hidden paper trail
The unusual request arrived in Haarlem, Netherlands, from a buyer identifying herself as Natalia from “2077AI.” She provided a spreadsheet containing 3,001 specific ISBNs—predominantly high-value academic texts published between 2020 and 2021 by institutions like Oxford University Press, Elsevier, and Wiley.
According to a recent Fortune report, de Vries and several other Dutch antiquarian booksellers dismissed the inquiry as fraudulent. But this massive, cross-border procurement effort highlights a growing industrial trend. Having largely exhausted the open internet for high-quality text, AI developers are now aggressively targeting the physical supply chain of global publishing to feed their large language models (LLMs).
The rise of “destructive scanning”
The fate of these books is a controversial practice known within the tech sector as “destructive scanning.” To rapidly digitize physical media, data brokers shear the spines off books, feed the loose pages through high-speed industrial scanners, and discard the physical remains.
This offline harvesting strategy was exposed last year during a legal battle involving Anthropic. Court documents revealed the company’s “Project Panama,” an initiative that purchased and destroyed millions of physical books to build a searchable training database.
While Anthropic maintains that its current acquisition programs do not target rare or antiquarian books, the destructive scanning practice itself was recently granted major legal cover. A federal judge ruled that utilizing legally purchased physical books to train AI constitutes fair use under copyright law, prompting Anthropic to settle separate, lingering claims regarding content downloaded from shadow libraries.
An emerging secondary market
The insatiable push for printed data has birthed a quiet but lucrative secondary market. Earlier this year, ISBNdb—a metadata cataloging firm—briefly advertised bulk book sourcing ranging from 1,000 to 1 million volumes, a service explicitly tailored to “LLM training needs” and “delivered at the scale AI demands.” While the company later claimed the webpages were merely an exploratory concept, similar bulk-buy anomalies have surfaced among secondhand booksellers in Germany and Switzerland.
For an industry built on ethereal algorithms and cloud computing, the mass eradication of printed academic literature serves as a stark reminder of its tangible—and increasingly aggressive—appetite for human knowledge.
