According to TechCrunch, citing 404 Media, a tracked rare book ended up at an Amazon facility in Las Vegas after being purchased through normal book channels. Amazon is acquiring hard-to-find titles, removing spines, and scanning them as AI companies hunt for fresh text to train large language models.
The easy AI training haul is exhausted. Public web text has been scraped to depletion, and older books now offer something the industry desperately needs: writing that predates the generative AI flood and won't contaminate models with synthetic data. Rare books are valuable precisely because they survived. Turning them into disposable training inputs marks a shift from scraping the open internet to consuming scarce physical works.
Amazon told 404 Media it buys books through commercial channels to improve customer products and services. That's legally defensible. It doesn't answer whether a tech giant should convert scarce physical works into one-time training inputs.
This is what the AI data shortage looks like in the real world: not abstract compute charts, but a warehouse knife through a spine. Amazon can claim it bought the books fair and square. Fine. But when the next generation of models depends on consuming material that was rare because it survived, the industry is telling you its appetite has outgrown the internet and is now reaching into the shelves. The economics are brutal and honest: if training data is the new oil, rare books are the last wells.
Filed to the Technology desk · Aug 17