Amazon Using Rare Book Destruction To Power Its Nova AI Models

When 404 Media planted an AirTag in a shipment of rare books, they found that Amazon has a process (VGT3) at one of its warehouses (Las Vegas) for ripping the spines out of books, scanning the pages and then destroying them — so it can train its Nova AI models. They are purchasing books published prior to 2022 with no digital version, which they refer to as “clean” content, meaning not produced by AI models.

Anthropic had a similar setup where they scanned books to train their AI model (“Project Panama”) but were able to make a legal case that the destruction of the physical books qualified as “fair use”. Amazon confirmed that they “purchase books through commercial channels to improve our products.” Physical books that don’t exist online are being turned into a proprietary corporate asset.