Used book sales in Japan are surging, fueled by what appears to be systematic purchasing by AI companies looking to acquire training data. Large shipments—sometimes reaching hundreds of tons—are being funneled into the country's logistics hubs at an unusual frequency and scale.
The buying pattern itself suggests AI training operations rather than typical resale. Historically, used bookstores in Japan moved primarily romance novels and manga. Now the demand has shifted dramatically toward philosophy, history, politics, medicine, and law—genres rich in structured information valuable for machine learning. This genre concentration is too deliberate to be coincidental.
The suspicion extends beyond acquisition: reports suggest these books are being destroyed after serving their purpose in model training. If confirmed, the practice raises questions about whether AI developers are systematically depleting rare and secondhand collections as a one-time training resource, then discarding them. The scale and selectivity of purchases make this industrial approach to book acquisition a notable—and potentially troubling—trend in how AI companies source training data.