Amazon Destroying Rare Books to Source AI Training Data

Amazon has been acquiring rare books in bulk, destroying their spines to scan them for artificial intelligence training purposes, according to an investigation by 404 Media. The company tracked a rare book to an Amazon facility in Las Vegas known as VGT3, which uses a dinosaur carrying a book as its emblem. Amazon stated it purchases books through commercial distribution channels to enhance its products and services. Large technology companies require massive quantities of text to train their language models, having already exhausted most freely available internet content. Rare and out-of-print texts represent a valuable new data source. Books published before 2022 hold particular significance since they were created before generative AI systems became prevalent. Using such pre-AI content helps prevent "model collapse," a phenomenon where model performance deteriorates when trained extensively on AI-generated text.