AI Companies Are Buying Up Old Books: The Last Training Data Free of AI Slop

AI labs have started buying old paper books in bulk to train models on them. Text published before the arrival of generative AI remains the last large data pool not contaminated by machine-made "slop." The data broker ISBNdb sells access to the catalog, and the labs themselves do not disclose who is funding the purchases, The Next Web reported in July 2026.

Why books free of AI slop are valuable

Old books are valuable because they are guaranteed to have been written by humans. Any text published before ChatGPT launched in November 2022 physically cannot contain neural-network-generated content. That "purity" is what turns library shelves into a scarce resource for model training.

The problem is how the data is extracted. To digitize a book with an industrial scanner, workers cut off its spine and run the pages through an automatic feeder, which physically destroys the book. According to The Next Web, the deals involve millions of volumes, and the name of the lab placing the order does not appear in them.

• ISBNdb is a data broker that sells labs access to a catalog of books
• November 2022 is the launch of ChatGPT, the dividing line between "clean" and contaminated data
• Digitization is destructive: spines are cut off and paper books are destroyed
• Millions of editions are in circulation
• Buyers do not disclose which AI company paid for the order

Why is the data running out?

The big labs have effectively exhausted the open internet as a source of quality text. As early as 2024, analysts at Epoch AI estimated that the supply of high-quality public human data could run out in the second half of the 2020s, and the race for the remaining "clean" corpora has only intensified since.

In that race, books are a premium asset. A book offers long, coherent, edited text with rich language, rather than fragmentary posts and comments, so a single volume carries the kind of quality labs are willing to pay brokers like ISBNdb for.

What is data contamination?

Data contamination is the situation in which training text scraped from the internet increasingly consists of the output of neural networks themselves rather than material written by humans. A model trained on its own "slop" gradually degrades, losing diversity and accuracy. Researchers described this effect in the journal Nature back in 2024 and called it model collapse.

Hence the hunt for physical books. The internet has been filling up with AI text at a rapid pace since 2022, and telling human writing from machine writing on the web is getting harder, whereas a book's publication date is a reliable marker that you are looking at pre-digital, human-written text.

"The world's best AI training data is sitting on a shelf," is how

The Next Web describes the data broker ISBNdb's pitch.

The legal side remains a gray area: buying and scanning other people's books to train commercial models is the subject of lawsuits from rights holders, and the buyers' anonymity only raises more questions about transparency.

What this means

For the industry, this is a signal of structural scarcity: high-quality human data is finite, and everything created after 2022 is valued less and less because of the admixture of AI content. As models demand ever more text, the "pre-digital" legacy of old books, archives and newspapers is becoming a strategic and physically exhaustible resource.