Bookworm · Mike Schmitz and Cory Hixson

228: AI Snake Oil by Arvid Narayanan & Sayash Kapoor

August 1, 2025·1 hr 36 min·5 clips

Getty Images sued us for $25,000 per blog post over stock photos an agency used years earlier.

1. Bookworm hosts Corey and Mike review AI Snake Oil by Arvind Narayanan and Sayash Kapoor. 2. Narayanan is a Princeton computer science professor and director of the Center for Information Technology Policy, and Kapoor is a Princeton PhD candidate who previously worked at Facebook. 3. The core question the hosts explore is whether the book's distinction between predictive AI (which it calls snake oil) and generative AI (which it mostly defends) holds up. 4. The hosts open with follow-up, including Corey's yellow-rated progress on forming a mastermind group and Mike's rundown of what energizes and drains him in his content work. 5. Corey introduces the book, noting its academic, citation-heavy, Princeton University Press style and that it was named one of Nature's top 10 books of 2024. 6. The introduction defines snake oil as anything that doesn't do what it's advertised to do, and argues predictive AI specifically fits that description. 7. Chapter two lists five reasons predictive AI fails, including that a good prediction can still lead to a bad decision and that people can strategically game opaque AI systems. 8. Mike highlights the book's finding that automated hiring tools can score candidates higher based on irrelevant background details like bookshelves or wall color, which he finds personally troubling given how many students struggle to get a human callback. 9. Chapter three covers the fragile families challenge, a large study that tried to predict children's outcomes using AI, prompting Corey to share his father's story of being capped by a school 'predictor' despite acing his classes. 10. The hosts discuss the book's claim that luck plays a larger role in success than most successful people would admit, and specifically more in success than in failure. 11. Chapter four turns to generative AI, which the authors treat far more favorably, while still raising concerns about training data sourced from scraping the internet. 12. Corey tells a personal story about a marketing agency's stock photos triggering a $25,000-per-post Getty Images copyright demand years later. 13. Mike says he'd celebrate the decline of the stock photo industry but acknowledges the separate, valid concern about generative AI training on copyrighted creative work without consent. 14. Mike repeatedly criticizes the book's academic writing style, saying information is technically all there but not memorably delivered compared to a trade nonfiction book. 15. Corey says the book's tables, like the 'landscape of AI snake oil, hype and harms' chart plotting benign-to-harmful against works-to-snake-oil, were among its most useful moments. 16. Both hosts agree the five reasons predictive AI fails are individually legitimate, even as they found the chapter's length repetitive. 17. The episode is an unscripted two-host conversation blending personal follow-up, career reflections, and chapter-by-chapter book breakdown. 18. The tone shifts between technical critique of the book's arguments and personal, sometimes emotional asides about fairness and bias. 19. Listeners interested in AI policy, hiring bias, and honest book critique would enjoy this. 20. Listeners wanting a purely enthusiastic AI book recommendation without academic-style critique might skip it.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Bookworm