Amazon Is Destroying Rare Books to Train AI

Amazon, whose history began with selling books, has found a new use for them. According to the outlet 404 Media, the company is buying up rare texts en masse, cutting off their spines and scanning the pages in order to use the contents to train language models.

To trace the journey of one such book, journalists slipped a tracker inside it. The device led them to an Amazon facility in Las Vegas known by the designation VGT3. The site's emblem is a dinosaur clutching a book in its claws. In a comment to 404 Media, Amazon stated that the company “acquires books through commercial channels to improve the products and services its customers use.”

Why paper books were needed

Training large language models requires colossal volumes of text, and the major players have by now all but exhausted the open resources of the internet. (Anthropic, as TechCrunch notes, at one point even went so far as to use pirated copies of books.) Rare editions — especially those that have long been out of print or are missing from the web entirely — are becoming a new and coveted source of data.

The special value of old books lies in the fact that texts published before 2022 were by definition written by people, not generated by a neural network. This matters because of what is known as “model collapse”: when a language model is trained on too large a volume of text produced by other models, the quality of its answers begins to degrade. Paper editions from decades past are guaranteed to be free of any such “contamination.”

The result is a peculiar paradox: a company that grew out of a love of reading and a trade in books now values them not for the content available to the reader, but for their suitability as raw material for machines — and at the cost of physically destroying the copies themselves.

Source: TechCrunch