Massive orders revive the used-book market and reveal the backstage of destructive scanning to train models
Listen to this article
Read by Anchor
The used-book market worldwide has seen a unusual purchasing pattern in recent months, as independent booksellers have recorded a flow of massive, exceptional orders that ship huge quantities of novels and old works to distant warehouses. At the “Barter Box” store in Northumberland, Stewart Manley was accustomed to selling between two thousand and three thousand books per week, yet a single consolidated order from a Canadian company equaled seven days’ worth of sales, which he described as something he had not witnessed in his thirty years in the trade. The case was not isolated; sellers in various locations reported similar, unprecedented orders that raise questions about their ultimate destination and the motive for buying them in such volume.
Growing suspicions point to AI companies seeking new information sources to train their large language models as the actual driver of this sudden surge.The chain traces back to a U.S. court ruling in 2025 in favor of Anthropic in a lawsuit brought by three authors, where Judge William Alsop held that using purchased books to train software constitutes a highly transformative use and does not violate U.S. copyright law. Recently declassified court documents revealed that training the chatbot “Claude” involved tearing and destroying real books while processing them.
Internal documents from Anthropic disclosed that the project to ingest old books was called “Project Panama” and aimed to conduct a destructive scan of all available books. The industrial technique removes the book’s spine and separates its cover to feed pages at very high speed through scanners for digitization, after which the paper remnants are recycled. A company spokesperson said the model is trained on a mix of publicly available web data, commercially purchased corpora, and self-generated data, confirming that buying books is a common practice in the industry and that data-acquisition programs do not purchase or destroy rare or antiquarian books.
The approach has sparked tangible concern among shop owners about the fate of the works. Although David Tobin, who runs “Valden Box” in north London, is comfortable selling titles that have been stacked for years, he expressed sadness at the prospect of their eventual destruction. Derek Walker, owner of “MacNautans” in Edinburgh, believes that destroying a copy of a recent academic text, of which only one hundred copies were printed and most are held in public libraries, may not constitute a major loss, but selling a single remaining rare eighteenth-century edition for destruction poses a serious cultural problem.
The variety of purchased titles, ranging from obscure Latin texts to cowboy novels, serves generative models’ need for rare and diverse texts that enrich their linguistic capabilities.Conversely, Professor Emily Hudson, an intellectual-property specialist at Oxford University, explained that UK law differs fundamentally from U.S. jurisprudence, requiring prior permission from the copyright holder to create a training library and to carry out the copying and training processes.