AI Developers Buy Used Books in Bulk for Training, Then Destroy Them
Anthropic
Google/DeepMind
AI developers are purchasing millions of used books through intermediaries to create high-quality training datasets, avoiding low-quality AI-generated web content. Companies like Anthropic have spent millions digitizing books and then destroying them, raising ethical concerns about the loss of rare editions and restricted access to scanned data.
AI developers are buying millions of used books via intermediaries to build high-quality training datasets for their models, avoiding the flood of low-quality AI-generated content on the internet. Anthropic spent several million dollars digitizing large numbers of printed books for training Claude and then destroyed the books, having purchased them from Better World Books. A court ruled the use was fair, but Anthropic faces a potential $1.5 billion fine over a library of 7 million pirated e-books; a similar lawsuit was filed against Google over Gemini training on unlawfully obtained book copies. The ISBNdb database, containing 111 million catalogued books, now offers wholesale purchasing services for AI developers, with orders ranging from thousands to a million copies per transaction. Tom Harvey, who worked on Project Panama at Anthropic, confirmed that digitization uses both non-destructive and destructive scanning methods, with the latter being cheaper and faster but destroying millions of books. This practice raises ethical concerns about removing valuable or rare books from circulation and restricting public access to scanned content.
- Сокращения
- ISBN = International Standard Book Number — Международный стандартный книжный номер
Source: 3DNews —
original
