Used bookshops sell out by the pallet: provenance is now a line item, not a research problem

Stuart Manley of Barter Books in Northumberland told the BBC that one recent bulk order from a Canadian company equalled what he would expect to sell in seven days, and that a typical week for him is two or three thousand books ^[1][2]. That puts a single purchase order in the range of a couple of thousand volumes ^[25], and the booksellers filling these orders say they do not know where the stock ends up ^[4].

Manley says he has "never seen the like of this after 30 years in the second-hand book trade" ^[5]. Booksellers elsewhere in the world report similar mega-orders ^[3]. David Tobin of Walden Books in north London has also had unusual sales, and told the BBC it is "very nice to sell some of these titles which haven't been sold for many years, but it would be sad if they are ultimately destroyed" ^[6].

The reason the trade suspects AI buyers is procedural rather than mystical. In 2025 a US judge ruled that using books purchased this way to train AI was not a copyright violation, with Judge William Alsup calling Anthropic's use "exceedingly transformative" in a suit brought by three authors ^[7][8]. Anthropic separately paid $1.5 billion to settle a class action over the use of illegal copies of books as training data ^[9]. Court documents unsealed last month showed books being destroyed in the course of training Claude, under an internal name, "Project Panama", with the stated aim to "destructively scan all the books in the world" ^[10][11][12]. Destructive scanning means shipping books to be digitised at industrial scale, cutting the spine off so pages can be fed through quickly, and recycling the remains ^[13].

An Anthropic spokesperson said Claude "is trained on a mix of publicly available web data, commercially acquired datasets, and data we generate ourselves", called book sourcing a widely used industry approach, and added that "none of our data acquisition programs buy and destroy rare or antiquarian books" ^[14][15][16]. Manley says the Panama name is new to him but the reality is not, that it has been discussed on bookseller forums, and that he does not know whether his books went to it or to a similar project at another firm ^[17].

Here is the part worth reading as a market signal. Gizmodo, citing 404 Media, argues that older books are becoming a premium product because anything published after the adoption of LLMs may contain LLM output, and that up to roughly 2022 only humans wrote books, which is the writing model builders want ^[19][20]. The purchasing behaviour is not neatly date-bounded on the evidence available: Manley describes the orders as having "no rhyme or reason", running from obscure Latin texts to cowboy novels, and experts told the BBC that the eclectic mix itself points to AI buyers looking for material not already in the pile ^[18][24].

The provenance cost is also jurisdictional. Professor Emily Hudson, an intellectual property specialist at Oxford University, told the BBC that the UK starting point is that building a training library and training on it both require the copyright owner's permission, unlike the US position ^[21]. The same pallet therefore carries a different legal bill depending on where the scanner sits ^[26].

Booksellers are drawing their own tiers. Derek Walker of McNaughtan's in Edinburgh notes that an academic text printed in 100 copies with 75 already in libraries is no great loss if one is destroyed, while he has sold books that are the only known surviving example of an 18th century edition ^[22][23].

What to watch: whether buyers begin specifying publication-date cutoffs in their orders, which would confirm the pre-2022 premium; whether the UK permission default pushes scanning work to US sites; and whether Anthropic's carve-out for rare and antiquarian books ^[16] is stated in a way any supplier can audit at the point of sale.