Tech firms buy old books to train AI, stirring legal concerns

Tech companies including Amazon, Meta and Anthropic are buying physical books as training material for AI models, with reports that rare and used volumes are scanned and then destroyed. The practice is aimed at sourcing edited, human-written text at a time when online material increasingly includes AI-generated content, which may be less useful for improving new models.

Mississippi bookseller David Hallinan said he received seven or eight unusual orders around May through AbeBooks, mostly for single volumes that cost no more than $25, and believes they may have been part of the buying surge. Other sellers on platforms such as Alibris, AbeBooks, Biblio and ISBNdb have compared similar orders tied to names including “Blue Finch Project,” “Green Parrot Project,” “Zoombooks” and “Scan PB.” Jackson’s Lemuria Book Store said it has not seen the same activity, partly because much of its inventory is not listed on third-party marketplaces.

Legal experts said the core issue is less about destroying purchased books than about retaining and potentially distributing digital copies. In the June 2025 Bartz v. Anthropic ruling, Judge William Alsup found that training generative AI on book material and building a research library can fall under fair use, while Mississippi law professors warned that existing law may struggle to address the speed and scale of AI systems consuming copyrighted works.