AI Companies Are Buying and Destroying Older Books at an Alarming Rate

As companies continue training large language models for artificial intelligence, they’ve begun to run into a problem. The issue? There’s too much AI slop in the world, especially online, and training new AI models on AI-generated text can cause issues. In particular, doing so can result in what’s known as “model collapse,” where newer models trained on data generated by older AI models end up producing worse models that are more prone to errors.

AI companies have found a solution, though. They’re buying up older books in bulk so their models can soak up all those words. And, in the process, they tear the books to shreds.

Silicon Valley is getting some help in this endeavor from companies such as ISBNdb. This company and its competitors offer the ability to place incredibly large bulk orders (sometimes up to a million books at a time), which can include books so rare that very few copies remain.

The worst part? Silicon Valley and these bulk ordering companies are aware of how bad this all looks. According to 404 Media, ISBNdb’s website stated, “The optics problem is real. ‘AI company destroys two million books’ is not a headline that generates sympathy.”

AI Companies and Literature

This isn’t the first time AI companies have been caught training AI models on books and then destroying them. Originally, companies that used books for training used digital books (many of them allegedly pirated). However, Anthropic was sued over the practice, and the courts ordered it to pay a $1.5 billion settlement to authors.

During that lawsuit, it came to light that Anthropic had an initiative called Project Panama that involved using millions of physical copies of books to train Claude, books that were then shredded. And, unfortunately for book-lovers, this practice is legal.

The legal concept of the first-sale doctrine allows purchasers to do whatever they want with the book they bought without anyone, including copyright holders, interfering. Hence, why books can be sold secondhand. A judge found that the way Anthropic was using books (scanning and destroying them) counted as “transformative,” making it fair use and fair that they weren’t paying authors.

The Use of Rare and Older Books in AI Training

Now, it seems as if these AI companies are turning to older and rarer books more often in an effort to train models on human language, rather than AI language. The Telegraph reported that antique book dealers in the Netherlands have seen a significant uptick in bulk orders, a rarity in that market. It’s so rare that one book dealer who received an email for 3,000 titles assumed it was a scam. Other dealers around the globe have also reported the same trend.

The safety of antique books seems uncertain. However, in the wake of widespread outrage, ISBNdb released a statement on its website denying it had done so. The statement read in part, “We’ve removed a recent landing page, ‘Printed Books Sourcing for Your AI LLMs Dataset Needs.’ The facts: ISBNdb has never purchased, scanned, or sold a book – for AI training or anything else. We don’t train AI models, and we never have. The page was a test of market interest; no such service was ever brought to life.”

Hopefully, the massive backlash to this practice will encourage other companies to back away from it, as well.

(feature image: Pexels/Sóc Năng Động)

Have a tip we should know? [email protected]