AI gives forgotten books a new lease of life

In a second-hand bookshop in Badalona, near Barcelona, Marcel Font observed something unusual. A Canadian company had started purchasing books that had been sitting on shelves for years, primarily non-fiction Catalan-language works. Simultaneously, Michael Strötter, a German bookseller, experienced a similar pattern, receiving orders from Zoom Books, a Canadian firm that specializes in selling and reusing second-hand books.

These orders appeared to be automated, with hundreds of books requested in quick succession. The companies behind these purchases have denied buying books specifically for AI training, but their refusal to disclose client identities has raised suspicions among booksellers. Anthropic, a developer of an AI language model, was involved in a copyright lawsuit where documents revealed they had purchased millions of books and employed contractors to dismantle, scan, and convert them into digital files for training purposes.

The legal question surrounding this practice is complex. While purchasing and destroying a physical copy may not inherently make it lawful for AI training under US law, fair use has been considered in specific cases. In Anthropic's case, a federal judge ruled that converting legally purchased books into internal digital copies could fall under fair use, as long as the resulting single digital copy was not distributed publicly.

However, this does not extend to unlawfully obtained books or covers the extensive practices of AI companies like Zoom Books. The US Copyright Office has cautioned against oversimplifying fair use to a single rule for AI training, urging a case-by-case assessment based on factors such as the nature of the works, the acquisition method, the purpose of use, and the impact on the original market.