AI companies quietly buying thousands of used books to train models

On a summer day in Haarlem, Netherlands, antiquarian bookseller Pieter de Vries received an email from a woman named Natalia, identifying herself as representing a company called "2077AI." The message requested what she called a "fairly large order" and included a spreadsheet with 3,001 ISBNs and instructions to match titles, prepare quotes, and estimate shipping costs to China. De Vries barely glanced at it, dismissing the message as spam or phishing without reading past the first lines. It was only weeks later, when a Dutch journalist investigating the procurement contacted him, that de Vries learned the request was connected to artificial intelligence. "I was shocked!" he told Fortune.

The 3,001 titles in the spreadsheet were predominantly academic works published between 2020 and 2021 by major publishers including Emerald Publishing, Elsevier, Wiley, Routledge and Oxford University Press, spanning subjects from business and education to engineering, public policy and medicine. De Vries was not alone in receiving the request—several other Dutch antiquarian booksellers got the same message and similarly dismissed it as a scam. For de Vries, who specializes in old and rare books and rarely receives bulk orders, the unusual nature of the request reinforced his skepticism. "These books are expensive and exclusive," he said. "So I do not do bulk trade."

The inquiry provided a rare window into how AI companies are sourcing training material beyond the open internet. Last summer, court records revealed that Anthropic had purchased millions of physical books, removed their bindings, scanned them, and discarded the originals to build a searchable digital library for training its AI models in what internal documents called "Project Panama," according to The Washington Post. The process, known as "destructive scanning," involves cutting the spine from a book so its pages can be fed through high-speed scanners before the remaining physical copy is thrown away. A federal judge later ruled that using legally purchased books to train AI models constituted fair use under copyright law, and the case settled. Anthropic stated that Claude models are trained on a mix of publicly available web data, commercially acquired datasets and internally generated data, and that it purchases books through regular commercial markets.

Earlier this year, 404 Media reported that ISBNdb—a company known for maintaining metadata tied to International Standard Book Numbers, the unique numerical identifiers found on most book barcodes—had begun advertising services to help AI labs source bulk printed book purchases ranging from 1,000 to 1 million books. According to the report, the company's website said the service was tailored to "LLM training needs" and "delivered at the scale AI demands." ISBNdb later told Fortune the service was never launched and that the webpages reflected only an exploratory concept. Similar reports have emerged in Germany and Switzerland, where local media reported that secondhand booksellers received unusual bulk orders for highly specialized books inconsistent with traditional collecting or resale. Zoom Books, a Canadian company named in some reports, denied the allegations and said the purchases were part of its regular recycling and trading model. Fortune contacted 2077AI for comment about the purpose of the procurement request and whether the books were intended for AI training, but the company did not respond.