Why AI companies are turning to physical books for data

Amazon is acquiring rare and out-of-print books to use as training material for its large language models — the complex systems that power tools like ChatGPT or Amazon's own AI assistants. To turn these physical objects into digital data, the books are systematically destroyed by having their spines cut off so the pages can be fed through high-speed scanners. This process transforms a physical archive into a mountain of digital text that the AI can then process, analyze, and learn from. The company has acknowledged that it acquires books through commercial channels to improve its digital products.

Training an AI is essentially a process of teaching it how human language works by showing it massive amounts of examples. The goal is to give the AI enough patterns to recognize so it can eventually predict the next word in a sentence accurately. So far, these companies have relied on the internet for this information, but they have largely run out of high-quality, human-written text. They are now hunting for fresh, reliable data sources that exist outside of the web. Older, rare books are highly prized because they were written long before AI existed. This is crucial because if an AI is trained on content created by another AI, its performance can suffer. This is known as model collapse, where the quality of an AI's output degrades over time, similar to making a photocopy of a photocopy until the image becomes blurry and unrecognizable. By using books published before the era of AI-generated content, companies hope to keep their systems sharp, accurate, and truly human-like.

This practice highlights a strange tension in the tech industry: as AI companies struggle to find new data to stay competitive, they are turning to the very physical artifacts that modern technology once rendered obsolete. For a company that built its fortune on selling books to now be destroying them to fuel a future of automation, it raises a quiet question about what we value more: preserving the physical history of our written word or using those pages as raw fuel for the next generation of software. As the supply of human-generated information online dries up, we should expect to see more of our physical world converted into fuel for these digital machines.