Amazon Is Destroying Rare Books to Train Its AI Models
An investigation by 404 Media has found that Amazon is buying rare books, cutting off their bindings, and scanning them to gather training data for its AI systems. The company confirmed it purchases books to improve its products and services.
A new investigation by 404 Media has revealed that Amazon is acquiring rare, hard-to-find books, physically removing their spines, and scanning their contents to build training data for large language models. Reporters traced the practice by planting a tracking device inside a rare book, which ultimately turned up at an Amazon facility in Las Vegas known as VGT3. The site is marked with a symbol depicting a dinosaur holding a book in its claws.
In a statement to 404 Media, Amazon said it "purchases books through commercial channels to improve the products and services customers use." The company did not provide further detail on the scale of the scanning operation or the fate of individual books.
Why rare books are valuable for AI training
Training large language models requires enormous volumes of text, and companies have already exhausted much of what's readily available online. Some AI developers, including Anthropic, have faced allegations of relying on pirated books to source training material.
Rare, out-of-print, or otherwise hard-to-access books represent a fresh, untapped source of data. Their value is heightened by the fact that anything published before 2022 is guaranteed not to have been generated by an AI system. That distinction matters because training LLMs on too much AI-generated text can trigger so-called "model collapse," a phenomenon in which an AI system's output quality deteriorates after ingesting excessive amounts of AI-produced content.
