A hidden destination for rare books has revealed how the race to build better AI may threaten irreplaceable pieces of literary history.
Amazon, which began as an online bookstore, is buying rare editions, cutting off their spines, and digitizing their pages to train artificial intelligence models. Journalists at 404 Media discovered this after placing a tracking device in a rare book that later arrived at an Amazon facility in Las Vegas.
The book was delivered to a center labeled VGT3. Its symbol is a dinosaur holding a book in its claws. In a comment to 404 Media, the company confirmed that it purchases books through commercial channels to improve its own products and services for customers.
Why Amazon Needs Rare Books to Train AI
Large language models require extremely large volumes of text. Technology companies have already used a significant portion of the materials available online to train artificial intelligence systems.
As a result, books that have not been digitized, are not publicly available, or have not been reprinted for a long time are becoming especially valuable. Rare editions may contain texts that are absent from online databases, digital libraries, and regular retail channels.
Why Editions Published Before 2022 Matter
Books published before 2022 have additional value for AI developers. Their content was guaranteed not to have been created by modern language models, so such texts can serve as material free from AI-generated content.
The active use of AI-written texts to train new models may lead to what is known as “model collapse.” This refers to the gradual decline in the quality of responses as a system increasingly trains on material previously generated by another artificial intelligence system.
Amazon’s practice raises concerns about the preservation of rare books: for rapid scanning, physical copies are effectively destroyed, even though they may have cultural, historical, and bibliographic value.