Amazon is reportedly acquiring numerous rare books, removing their spines, and digitizing them for artificial intelligence model training, as detailed by 404 Media. The publication utilized a tracking device in a rare book, which ultimately led to an Amazon facility in Las Vegas.
This particular facility, identified as VGT3, is marked by an emblem depicting a dinosaur clutching a book. In response to 404 Media’s inquiry, Amazon stated that it “purchases books through commercial channels to improve the products and services customers use.”
Companies like Amazon require enormous quantities of text to train their large language models (LLMs). Having already consumed vast amounts of data from the internet (and, in Anthropic’s instance, through unauthorized acquisition of books), rare texts, especially those that are out of print or not available online, represent a valuable new source of training material.
These older publications are particularly valuable because there is no possibility that content published before 2022 was generated by an LLM. Training LLMs on AI-generated text carries the risk of “model collapse,” a phenomenon where the quality of an LLM’s outputs deteriorates from ingesting excessive amounts of AI-produced content.