Anthropic purchased millions of physical books and used destructive scanning to create training data for its Claude AI chatbot [1].
The practice highlights the aggressive pursuit of high-quality text data by AI developers as they exhaust available digital archives. The method used by Anthropic, removing bindings to facilitate rapid scanning, has sparked debate over the preservation of physical media.
According to documents made public in a lawsuit filed earlier this year, the company hired contractors to slice the bindings off the books [2, 3]. This process allowed the pages to be scanned more efficiently, though it rendered the original physical copies unusable [2]. The company sought this large-scale text data to improve the performance and knowledge base of the Claude model [1, 3].
While some reports suggested the books were being burned, other investigations found this characterization to be misleading. The documents indicate the books were scanned rather than incinerated [3].
Anna Rascouët‑Paz of Snopes said, "I examined hundreds of pages of the company's internal documents" [3]. Those documents confirm the scale of the operation involved millions of books [2].
The activity took place earlier in 2024 [2, 3]. Anthropic has not provided a detailed public response to the specific methods used by its contractors during that period.
“Anthropic purchased millions of physical books and used destructive scanning to create training data”
This reveals a shift in AI development where companies are moving beyond web-scraping to acquire physical assets for digitization. By utilizing 'destructive scanning,' AI firms can process vast amounts of printed knowledge faster than traditional archival methods, though it creates a permanent loss of the physical artifacts in exchange for digital training weights.
