TL;DR
Steve Matthews, who runs Bookcase in Carlisle, has spent two months receiving orders running from half a dozen up to twelve volumes, with no thematic connection between them — a pattern he attributes to AI training data collection rather than any reader. The shop’s online catalogue runs somewhere in the 30,000 to 40,000 range. Booksellers in Britain and Ireland report the same thing, as do sellers in North America, Australia and mainland Europe.
What makes the orders look wrong
The tell is incoherence. A collector buys within a subject; these orders span unrelated and often deeply obscure titles. Matthews’ reading of it is blunt: somebody is working through every book that exists, one shop at a time.
The suspicion has documentary support. US court filings revealed Anthropic had bought physical books in bulk for Claude’s training under a programme called Project Panama, which used destructive scanning — books taken apart so the pages can be digitised at speed. Matthews assumes the volumes leaving his shop meet the same end. Anthropic’s response is that “none of our data acquisition programs buy and destroy rare or antiquarian books” — a denial worth reading precisely, since it covers rare and antiquarian stock and says nothing about ordinary second-hand copies.
The legal position is not settled here
Adam Newell of Withnail Books in Penrith argues that buying in-copyright books to train a model is plainly unlawful and hopes the ongoing lawsuits succeed. Matthews is more measured, granting AI’s usefulness for organising knowledge while objecting that ingesting a whole text and reusing it without credit is a larger infringement than the usual copyright argument contemplates.
A 2025 US ruling found Anthropic’s use of lawfully purchased books “exceedingly transformative” and therefore permitted under American fair use. That doctrine has no direct UK equivalent, and the government’s text and data mining consultation left the question unresolved — so the ruling settles less for British rightsholders than the headlines implied.
Looking forward
Newell wants a legal requirement for human-authored labelling, noting some publishers already print a no-AI mark on covers. That would be a significant intervention in UK publishing.
The immediate point is quieter. Two days ago Southampton dropped Turnitin over training data concerns — an institution with lawyers acting on principle. This is the same fight reaching people with no leverage at all, who can only sell the book and guess what happens to it.