Amazon destroys rare books to source AI training data

Amazon's practice of sourcing rare books for AI training reflects a broader industry challenge: the finite supply of authentic human-written text online. As LLMs (AI models like ChatGPT) have saturated freely available internet content, companies have begun seeking alternative sources—Anthropic, for example, has already used illegally pirated books. The stakes are concrete: when LLMs train on too much AI-generated text, a phenomenon called 'model collapse' degrades their output quality, making pre-2022 human-written texts especially valuable.

Rare and out-of-print books are a particularly coveted resource because they are nearly impossible to find on the internet, yet contain guaranteed human authorship. Amazon's centralized scanning operation at its VGT3 facility in Las Vegas—identifiable by its dinosaur-holding-a-book symbol—suggests the company is investing infrastructure in this acquisition strategy. The fact that Amazon publicly frames this as purchasing books 'through commercial channels' indicates that such sourcing, while potentially destructive to irreplaceable cultural artifacts, is now an accepted part of AI development economics.