In a modest second-hand bookshop in Badalona, near Barcelona, Marcel Font began to notice something he had never seen before. A Canadian company had begun buying books that had languished on shelves for years, many of them non-fiction Catalan-language works. The orders sometimes came only minutes apart.
Hundreds of kilometres away, the German bookseller Michael Strötter encountered a strikingly similar pattern. At 2:53 a.m. on 30 April 2026, the first order arrived. More followed throughout the night. In almost every case, the buyer was the same company: Zoom Books, a Canadian firm that describes its business as selling and reusing second-hand books, while recycling those it cannot sell.
The books being sought were neither rare first editions nor costly manuscripts. They were cookbooks, novels, biographies, and ageing specialist titles. Similar accounts in the UK, Australia, New Zealand, and elsewhere emerged, all describing the same recurring pattern: large orders, at times seemingly automated, spanning an extraordinary range of titles and often requesting only a single copy.
These companies have denied buying the books for AI training purposes, but their refusal to reveal the identity of some clients, citing confidentiality agreements, has fuelled suspicion of ulterior motives.
In a copyright lawsuit against Anthropic, the developer of Claude, court documents revealed the company had purchased millions of printed books and then hired contractors to remove their covers, cut away their pages, scan them, and convert them into digital files before disposing of the physical copies.
The undertaking, known as ‘Project Panama’, frequently relied on ‘destructive scanning’, a process in which a book is dismantled so that its pages can be digitised more rapidly. The concerns voiced by booksellers therefore have a known precedent: acquiring books that are difficult to find in digital form and converting their contents into data that can be used to train language models.
The legal question is considerably more complex than the formula ‘buy the book, then destroy it’. Under US law, purchasing a physical copy and destroying it does not, by itself, make its use in AI training lawful.
In the Anthropic case, federal judge William Alsup ruled in June 2025 that converting lawfully purchased books into internal digital copies could fall within fair use. His reasoning rested partly on replacing each physical copy with a single digital copy that was not distributed to the public. He also found that using the books to train the model qualified as fair use in the circumstances of the case.
The court nevertheless distinguished those books from the millions of copies Anthropic had obtained from piracy sites. Their unlawful acquisition did not become legitimate simply because the material was subsequently used for training.
Destroying the physical copy was therefore one factor in the court’s legal analysis, rather than a general rule. Nor does the $1.5bn settlement approved in July this year over pirated books confer on AI companies an unrestricted right to use any book merely because they have purchased a copy.
The US Copyright Office has cautioned against reducing ‘fair use’ to a single rule applicable to every instance of AI training. The assessment depends on the nature of the works involved, the manner in which they were obtained, the purpose for which they are used, and the effect of the resulting outputs on the original market
