🚨 New paper on AI and Copyright 🚨 It's easy to dismiss AI-written books 📚 as slop nobody buys. So, armed with full-text AI detection from
- How we did it 📊: We randomly sampled 14,419 self-published genre fiction ebooks (less gatekeeping and low barrier to entry) and ran Pangram on the full-text to classify each book as substantial AI (>25%), light AI (<25%) and no-AI(0%). Then we matched each book to aBooks with substantial AI text are now everywhere: 1 in 5 books in our sample, across every genre - romance, thrillers, sci-fi, all of it. The typical book written with AI sells far fewer copies than the typical non-AI one. But there are so many of them that as a group they'veHere's the imbalance at the heart of the paper. Over 3 years, the books catalog grew 38x cumulatively. The book's revenue? Only 9x quarterly. Way more books fighting over a pie that barely grew which means the average book earns less than it used to. (4/n)"Sure, but doesn’t it mean that cheap AI books are dragging the average down?" No, and this is our key result. No-AI books, on their own, earn less per book than they did in 2023 in 7 of 8 genres. The one genre where human authors are doing better (+35%) is Fantasy/horror whichThe more AI books enter a genre, the more ground non-AI books lose. Their hold on the top-chart spots slides from ~88% in the least-AI genres to ~63% in the most. And it's the worst in Kindle Unlimited-heavy genres since KU readers draw from one shared subscription pool, so everyThe million dollar question : Who's producing these books? Authors who, once they started producing with AI, didn't stop but instead accelerated. Among those who kept at it, 3 in 4 published faster than before, some hitting dozens of books a year. The monetary earnings areAnd the AI books that succeed share a signature. They contain noticeably more rare phrases that already appear in published books than successful non-AI books do.Award-winning literary fiction, for comparison, reuses rare phrasing at less than half their rate. (8/n)What’s funny is that AI is causing mode collapse :) A lot of AI-written books have verbatim overlap from other AI books. Can this serve as an alternative evidence to AI-detection skeptics ? (9/n)👨⚖️ TL;DR: AI doesn't need to write high-quality books for non-AI book writers to get hurt. Cheap books at massive volume are enough and they've cut what the average human-written book earns in most genres and pushed human titles down the bestseller ranks. (10/n)Our results also speak directly to a copyright question that is being litigated right now: whether AI outputs compete with the very works the models are trained on. In Kadrey v. @Meta, Judge VinceChhabria spoke about market dilution and flagged this as the harm that could beOur paper can be found at the link below :arxiv.org/pdf/2607.20349Anonymized data and code coming soon. Stay Tuned. Many many thanks to friends at@pangram@max_spero_@kthai1618@bradley_emi@jennajrussellfor their encouragement and support with this research 💕 # Join the conversation
source & further reading
twitter.com — original article