AI Companies Are Bulk-Buying Secondhand Books — The Strange Gold Rush for “Unpolluted” Training Data

Introduction: The Strange Orders Landing at Bookshops

On August 15, 2026, The Guardian reported that secondhand booksellers across the UK and Ireland are being hit by a wave of bulk orders from mystery buyers — and the working theory is that AI companies are buying the world’s used books to feed their models.

The orders are bizarre. One recent request included an Estonian translation of John le Carré’s The Mission Song, a copy of Anne Brontë’s Agnes Grey from a specific imprint, and the October 1983 edition of Warship, a monthly magazine. “Normally a book order is on a theme but this is all over the place,” Stuart Manley, co-owner of the famous Barter Books in Alnwick, Northumberland, told the paper. His theory about who’s paying: “masses and masses of AI money being squandered.”

If that theory is right, the story is bigger than quirky bookshop gossip. It’s the clearest physical-world sign yet that the AI industry has hit its data wall — the point where all the easily-scraped internet text has been consumed, and companies are turning to printed books published decades before ChatGPT existed: text that is genuinely fresh to a model and untouched by AI-generated slop.

Anatomy of the Buying Spree

The Guardian spoke to sellers across the UK and Ireland, and reports from the US, Australia, and continental Europe describe the same pattern: large, thematically random orders placed through marketplaces like Biblio and the Amazon-owned AbeBooks, arriving from buyers in the US, Canada, and Europe.

Several sellers flagged the same fingerprints: opaque buyer aliases, orders from different accounts shipping to the same address, and zero interest in the discounts any normal trade buyer would demand. One buyer cited by multiple sellers, the Canada-based Zoom Books, describes itself as “North America’s leading book recycling company.” As one UK seller put it: “It’s disrupting the secondhand bookselling business because this isn’t typical for the trade.”

The Anthropic Precedent: Spines Sliced, Scanned, Recycled

The booksellers’ suspicions have a concrete anchor. The Washington Post revealed this year that Anthropic — the company behind Claude — spent tens of millions of dollars acquiring books, slicing off their spines so the contents could be scanned, and then sending the books for recycling, all under a program literally labeled “data acquisition.”

Anthropic’s response to the Guardian is a rare on-record glimpse into how frontier labs source text: “Claude is trained on a mix of publicly available web data, commercially acquired datasets, and data we generate ourselves. Sourcing books is a widely used approach for training large language models across the AI industry. None of our data acquisition programs buy and destroy rare or antiquarian books.”

In other words: buying books at scale is standard practice — and Anthropic says it is not alone. Last month, 404 Media reported that the books database ISBNdb was telling booksellers “the world’s best AI training data is sitting on a shelf,” arguing pre-2022 books are ideal because their text is unpolluted by chatbot-generated material. (ISBNdb later called that page a “test of market interest” and took it down.)

Why Books? The “Unpolluted” Data Premium

Why would a lab pay “top whack” for a 1983 naval magazine? Two words: data quality. Models are trained on vast amounts of text, and the pool of high-quality human writing is finite. Worse, everything published online since late 2022 is increasingly likely to contain AI-generated content — and training new models on model-written text risks model collapse, the degradation loop where synthetic output crowds out genuine human signal.

Pre-2022 printed books solve both problems at once. They are human-written by definition, professionally edited, frequently absent from the open web, and legally cleaner to acquire than scraping — buying a physical copy through normal channels looks more like licensing than the mass scraping that has dragged AI companies into copyright court. The result is a strange new commodity market where your grandparents’ bookshelves are a premium data mine.

Hitting the Data Wall

The used-book gold rush is one front in a broader scramble: labs repricing APIs as demand strains supply, publishers cutting licensing deals, and startups building distillation and synthetic-data pipelines to squeeze more from less. Frontier labs now treat proprietary data — books, archives, licenses, human feedback — as a strategic moat, much like chips and power. Data provenance is becoming part of a product’s spec sheet, not a footnote.

What This Means for AI Tool Buyers in 2026

• Treat data provenance as a buying criterion. Vendors that license or legitimately acquire their training data face fewer copyright shocks — and less risk that your workflow gets disrupted by a court order. Ask what the model was trained on before you build on it.
• “Clean” data is a quality signal. Models trained on curated, human-written corpora (books, licensed archives) tend to write better prose and hallucinate less than those retrained on recycled web text. It’s a fair question to put to any AI writing or research tool.
• Expect licensing to become a feature. Just as music AI tools began advertising label deals, expect chatbots and writing tools to advertise “licensed book corpora” the way they tout benchmark scores — Anthropic’s own statement reads like the start of that messaging.
• Content creators have new leverage. If pre-2022 text is this valuable, rights holders of books, archives, and back catalogs are sitting on negotiable assets — and tools that pay creators for training data will say so.

Want to compare AI tools that publish their data practices? Browse the vetted catalog on aitrove.ai.

Frequently Asked Questions