Anthropic used specialized system to scan Japanese books for AI training

Court documents from a lawsuit against Anthropic have revealed that the AI developer utilized a specialized system, dubbed ‘Nippan Loader’, to acquire and digitize Japanese books for training its Claude models. Internal records indicate that Anthropic contracted with Nippan, Japan’s largest book wholesaler, to purchase large quantities of physical books for shredding and scanning.

Parallel reports from Japan highlight a surge in orders from various online secondhand bookstores, with many shipments being routed to a single logistics center in Okayama Prefecture. These orders often consist of niche professional texts in fields such as philosophy, medicine, and law. Trade data suggests that over 50 tons of Japanese books have been exported to the United States since last year, potentially linked to AI data procurement efforts.

The Japan Book Publishers Association has formally requested an explanation from Nippan regarding its transactions with Anthropic. The situation has sparked debate over the boundaries of Japanese copyright law, specifically regarding whether large-scale scraping and physical destruction of books for AI training constitutes ‘unreasonable harm’ to copyright holders.