Researchers sampled 14,419 ebooks and matched them to three years of sales data
TuhinChakr · x · 2026-07-28
- The study randomly sampled 14,419 self-published genre-fiction ebooks, then used Pangram to classify each full text into substantial AI (>25%), light AI (<25%), or no AI (0%).
- Each book was matched against a proprietary daily sales feed from a Big Five publisher, giving the researchers book-level AI-content and sales trajectories for three years.
- The authors frame the paper as evidence for a live copyright question: whether AI outputs compete with the works the models were trained on.
Related event: Nearly a Third of Bestselling New Books Show AI Traces(2 posts)→
More from Safety
- NVIDIA expands Open Secure AI Alliance with companies focused on AI security — nvidia · 2026-07-29
- After the Hugging Face hack, one AI safety critic says scalable sandbox research is still missing — basedjensen · 2026-07-29
- Sam Altman calls AI sandbox breakout a security and alignment failure — victor_explore · 2026-07-29
- Stanford HAI warns world models need a new governance agenda before safety-critical deployment — StanfordHAI · 2026-07-29
- The AI Governance Dilemma: Why Human Steering Can't Outpace System Evolution — anderssandberg · 2026-07-28
- Young adults use chatbots to rehearse social interactions as EU AI Act enters its main phase on Aug. 2, 2026 — emmanuelvivier · 2026-07-28