Amazon reportedly scans dismantled rare books for AI training
Amazon is reportedly buying rare books, removing their spines and scanning them to create training material for artificial intelligence systems. In an investigation by 404 Media, a tracking device placed in a rare book led to VGT3, an Amazon facility in Las Vegas. The site identifies itself with a logo depicting a dinosaur holding a book in its claws.
Amazon told 404 Media that it “purchases books through commercial channels to improve the products and services customers use.” The report does not specify the AI model or product for which the scanned texts are intended, but it describes a physical supply chain for acquiring and digitising printed material.
Rare texts become a source of training data
Large language models require vast volumes of text. Much of the material readily available online has already been collected for model development, making books that are out of print or otherwise unavailable on the web a potentially valuable source of additional text.
Older publications also have a distinct property: anything published before 2022 predates the widespread release of generative AI tools. That matters because models trained on too much AI-generated material can face “model collapse”, a decline in the quality of their output after ingesting synthetic text.
The report places Amazon’s reported book acquisition activity alongside its much broader AI investment. The company’s relationship with Amazon’s $5 billion Anthropic investment includes Amazon’s $5 billion commitment and a stated plan to spend $100 billion on AWS, illustrating the scale at which AI infrastructure and model development are being pursued.
Physical digitisation raises operational questions
Removing a book’s spine enables high-volume scanning, but it also destroys the original copy. The report highlights a tension between the demand for high-quality, human-produced training data and the preservation of rare physical works, particularly where those texts cannot easily be found online.
Amazon’s statement frames the purchases as commercial transactions intended to improve customer-facing products and services. It does not address the handling of individual rare volumes, the selection process for titles, or whether alternative non-destructive digitisation methods are used.
For organisations developing, buying or governing AI systems, the practical implication is to assess training-data provenance alongside model performance: acquisition channels, rights, preservation impacts and the risk of synthetic-data contamination can all affect how AI programmes are evaluated and managed.