AI Training Demand May Be Driving a Secondhand Book Boom

Independent booksellers are reporting an unexpected wave of bulk purchases involving titles that may have sat unsold for years. The orders are often large, international and difficult to explain through ordinary reader demand, leading sellers to suspect that physical books are being acquired for artificial-intelligence training.

The final destination has not been confirmed in many cases. Booksellers know where packages are shipped, but not necessarily who scans the books or how the resulting text is used. The AI explanation is therefore a credible hypothesis supported by industry practices, not a proven account of every order.

Orders unlike anything sellers have seen

Barter Books in Northumberland normally sells between 2,000 and 3,000 books in a week. Owner Stuart Manley said one recent order from a Canadian company equalled a full week's normal sales. After three decades in the trade, he had not seen a similar pattern.

Other sellers have reported orders spanning unrelated subjects rather than the consistent choices expected from a specialist collector, library or retailer. That randomness may be useful to an AI company seeking text that is absent from common web datasets.

For booksellers, the sudden demand has an obvious benefit. Slow-moving stock generates revenue and frees shelf space. The unease concerns what happens after the sale.

Destructive scanning turns paper into data

Industrial digitisation often removes a book's spine so pages can be fed quickly through a scanner. The physical remains are then recycled. The method is efficient, but the book cannot be restored.

Court documents connected to Anthropic revealed an internal effort known as Project Panama, which involved buying and destructively scanning books for training Claude. Anthropic has said its models use publicly available web material, commercially acquired datasets and internally generated data. It also said buying and scanning books is used across the AI industry and that its programmes do not target rare or antiquarian works for destruction.

The disclosure helps explain why bulk buyers might value physical copies even when digital editions exist. Purchased books can provide a documented acquisition trail and material that is not readily available online.

Copyright law differs across borders

In a 2025 US case brought by authors, Judge William Alsup ruled that using lawfully purchased books to train AI was highly transformative and permitted under US fair-use law. That ruling did not create one worldwide rule for training data.

Oxford intellectual-property specialist Emily Hudson told the BBC that the UK starts from a different position: copying books to build a training library and conduct training generally requires permission from the copyright owner.

That difference creates risk when books are purchased in one country, shipped to another and processed for a model used globally. Ownership of a physical copy does not automatically grant every right to reproduce its contents.

Rare titles create a preservation problem

Many secondhand books exist in enormous numbers. Recycling one unwanted copy may have little cultural cost. The problem is that rarity is not always obvious to a bulk purchaser.

A modest academic work may have had only 100 copies printed. An old edition may be the only surviving example outside a private collection. If those books are destroyed without bibliographic checking, the data survives while the physical object and its historical features are lost.

Booksellers can reduce that risk by screening unusually scarce items, checking catalogues and excluding antiquarian material from bulk sales. Buyers can also adopt policies requiring non-destructive scanning for rare editions.

A new market with unanswered questions

The AI training secondhand books boom shows how demand for data can reach into unexpected parts of the economy. A technology company may see millions of pages of useful language; a bookseller sees stock, and a historian sees physical evidence worth preserving.

Transparency would make those interests easier to balance. Bulk buyers should identify the intended use, explain whether books will be destroyed and publish safeguards for rare material. AI developers should also document licensing and jurisdiction rather than assuming one US judgment settles every copyright question.

For now, the trade is enjoying sales while wondering who the true reader is. The answer may not be a person turning pages, but a model absorbing text at industrial scale.

This Novexa News technology report is based on bookseller accounts, court records and reporting by the BBC.