AI developers, including Anthropic, are purchasing physical books in bulk to scan for training data and then destroying the original copies.
This practice highlights the growing desperation of large-language-model developers to find new text sources as they exhaust the supply of publicly available digital data.
Reports from 2026 indicate that some companies are buying and destroying books in batches of "tens of thousands" [1]. One Dutch bookseller reported receiving a single order for 3,001 books specifically for AI training [2]. These operations are designed to convert physical knowledge into digital tokens for machine learning.
An anonymous bookseller said, "We are buying books in bulk, scanning every page, and then discarding the physical copies once digitised" [3]. Some reports suggest these companies use middleman services to purchase the volumes anonymously to avoid public scrutiny [4].
However, the scale and nature of the destruction are disputed. An opinion piece in The Globe and Mail said, "Anthropic has purchased – and destroyed – millions of books in pursuit of their content" [1]. This claim of mass destruction is contested by other investigators. The Snopes Fact-Check Team said, "There is no verifiable evidence that AI firms are systematically destroying rare books; the claim is unsubstantiated" [5].
Internal documents from Anthropic, referred to as Project Panama, reportedly describe a coordinated effort to acquire this data [1]. While the companies seek the information contained within the pages, the physical disposal of the books has sparked a debate over the preservation of physical media in the age of artificial intelligence.
“Anthropic has purchased – and destroyed – millions of books in pursuit of their content.”
The transition from scraping the open web to purchasing physical archives signals a critical inflection point in AI development. As 'data exhaustion' becomes a reality, AI labs are moving toward high-quality, curated human knowledge found in print. This creates a tension between the efficiency of digital scaling and the preservation of physical cultural artifacts, while raising questions about the ethics of treating books as disposable raw material for software.
