Something unusual is happening in the second-hand book trade. Independent booksellers across Britain and beyond have reported a sudden wave of bulk orders, with large quantities of old and obscure titles being shipped to distant warehouses for reasons that remain unclear.<\/p>\n
At Barter Books in Northumberland, owner Stuart Manley typically sells between two and three thousand books in an average week. Recently, however, a single order from a Canadian company matched roughly a full week\u2019s normal sales. After three decades in the trade, Manley said he had never seen purchasing activity on this scale.<\/p>\n
Other booksellers have noticed similar patterns. The orders often appear random, ranging from obscure Latin works to cowboy fiction, and sellers say there is little obvious connection between the titles being bought.<\/p>\n
The leading suspicion is that artificial intelligence companies, or firms working on their behalf, are acquiring physical books as fresh training material for large language models.<\/p>\n
The theory gained momentum after a major US copyright ruling involving Anthropic, the company behind the Claude chatbot. In 2025, a federal judge ruled that using lawfully purchased books to train an AI system could qualify as fair use under American copyright law.<\/p>\n
Second-Hand Books are Selling Faster Than Ever and AI May Be One Reason (Picture Credit – Pinterest)<\/p>\n
The lawsuit had been brought by three authors, and Judge William Alsup described Anthropic\u2019s use of the books as highly transformative. That decision drew a significant distinction between books that had been legally acquired and material obtained through unauthorised sources.<\/p>\n
When court documents were later made public, they revealed another controversial detail. Some physical books had been destroyed after being digitised for AI training.<\/p>\n
Anthropic said Claude is trained using a mixture of publicly available information, commercially acquired datasets and material generated internally. The company also said the acquisition of books for AI training is not unusual within the industry and maintained that it does not purchase and destroy rare or antiquarian volumes.<\/p>\n
Internal Anthropic communications reportedly referred to the book digitisation effort as \u201cProject Panama\u201d. The documents suggested an ambition to scan books at enormous scale using a process known as destructive scanning.<\/p>\n
That method involves removing the spine of a book so individual pages can be fed quickly through industrial scanners. Once digitised, the remaining paper can be recycled.<\/p>\n
For second-hand booksellers, the idea is unsettling because the ultimate destination of the books is usually unknown. Manley said the subject has become a regular topic of discussion among dealers, even though there is no proof that the recent orders received by his shop are linked to Anthropic or any other AI company.<\/p>\n
The unusual mix of titles has only deepened the speculation. Researchers have noted that rare, specialised or less digitised books could be particularly useful to AI developers because they offer information and language patterns that may not already exist in widely scraped online datasets.<\/p>\n
The legal position is also more complicated outside the United States. Oxford University intellectual property specialist Professor Emily Hudson has noted that UK copyright law does not automatically provide the same freedom to copy books for AI training.<\/p>\n
Under British law, making copies to create a training dataset can require permission from the copyright holder. That means practices allowed under certain US interpretations of fair use may not transfer neatly to the UK.<\/p>\n