A US judge cleared Anthropic's purchase and destruction of used books to train Claude as fair use. Booksellers say the bulk buying boom is already at their counters.
For three decades, Stuart Manley has sold a few thousand used books a week from his shop, Barter Books, in Northumberland. Then a single order from a Canadian company arrived that equalled a full week of sales. He has worked the secondhand trade for 30 years and told the BBC he has never seen anything like it.
Barter Books is not alone. Independent sellers from the UK to the wider English-language secondhand market are reporting the same pattern: sudden, irregular mega-orders that bear no resemblance to the slow, idiosyncratic trade of remaindered and antiquarian stock. The orders are not aimed at overseas readers. The titles range from obscure Latin texts to cowboy paperbacks, and the volume of any given title often has no relation to its cultural cachet.
Booksellers' best guess, voiced on trade forums and to reporters, is that the books are bound for AI training pipelines rather than resale. That suspicion is not yet confirmed for the Canadian buyer behind Manley's order. The only documented link between the secondhand boom and an AI lab runs through a US courtroom, a 2025 ruling against Anthropic, and a stack of unsealed court documents.
The ruling came in a lawsuit brought by three authors against the AI company behind the Claude chatbot. US District Judge William Alsup held that Anthropic's use of their books to train its model was "exceedingly transformative" — a high bar the judge said the use cleared — and therefore not a US copyright violation, according to the BBC's account of the case. The decision set a permissive precedent for an industry-wide practice: buying secondhand books in bulk, then ingesting them as training data.
What the public did not learn until last month was what happened to the books after they were bought. When the court documents were unsealed, they revealed that Anthropic's internal codename for the book-ingestion effort was "Project Panama," and that its stated aim was to physically destroy and scan purchased books at scale. Destructive scanning is the industrial process of shipping books to a scanning facility, removing the spine so every page can be captured at speed, and recycling what remains.
An Anthropic spokesperson, quoted by the BBC, said Claude is "trained on a mix of publicly available web data, commercially acquired datasets, and data we generate ourselves," and described sourcing books for training as "a widely used approach across the AI industry." The company added that "none of our data acquisition programs buy and destroy rare or antiquarian books."
That last claim is the one booksellers say is hardest to verify at the till. Manley told the BBC the orders he sees appear random, with "no rhyme or reason," and include titles unlikely to be sought by human collectors. David Tobin, who runs Walden Books in north London, said the sales are welcome for stock that has sat unsold for years, but added that it would be sad if rare books were ultimately destroyed after surviving this long.
The legal landscape outside the US is different. Emily Hudson, an intellectual-property specialist at Oxford University, told the BBC that the UK starting point is the opposite of the US fair-use framework: copying to build a training library and copying again to run the training both require the copyright owner's permission. The Anthropic precedent is therefore a US judgment, not a global one, and does not by itself authorise the same pipeline in London or Toronto.
What it does authorise, in practice, is the pattern booksellers are already seeing. The chain from a Barter Books shelf to a Claude training run now has a name, a project code, and a court ruling behind it. The chain from Manley's Canadian buyer to any specific lab does not, and Anthropic's spokesperson said its own programs do not target rare or antiquarian titles.
For the trade, the open question is which links in that chain become routine. Derek Walker, who runs Edinburgh's McNaughtan's, framed the dilemma plainly: a contemporary reprint sold in bulk for destruction is one thing; the only known surviving copy of an 18th-century edition is another. The secondhand market can absorb the first. It cannot tell, on any given order, which one it is shipping out the door.