AI Snake Oil

**Author:** Arvind Narayanan **Estimated Reading Time:** 45 minutes

You will learn how to distinguish between AI that genuinely works and AI that merely claims to work. You will understand why predictive AI so often fails when applied to human lives, what generative AI can and cannot do, why existential risk debates distract from real harms, and how institutions can respond to AI without falling for hype.

This book is for anyone who has felt uncertain about AI claims. It is for professionals evaluating AI tools, policymakers shaping technology rules, educators navigating generative AI in classrooms, and citizens trying to understand what artificial intelligence means for their lives. You do not need a technical background. You need curiosity and a willingness to question what you are being sold.

Artificial intelligence has become a cultural obsession. Every week brings announcements of breakthroughs, warnings of doom, and promises that AI will transform medicine, education, business, and art. The language around AI has grown so inflated that it has become difficult to know what is real. Some claims are true. Many are exaggerated. A significant number are simply false. The term "AI snake oil" describes a specific phenomenon: AI that does not and cannot work as advertised. This is not a matter of technology being immature or needing more data. Snake oil AI fails because the underlying assumptions are wrong. The problem is not that the algorithms need improvement. The problem is that the entire approach rests on a flawed understanding of what AI can do. Arvind Narayanan, a computer scientist at Princeton University, has spent years studying how AI actually functions when it meets the real world. His research spans privacy, security, and the social impact of technology. What he has found is troubling. AI systems are being deployed in high-stakes settings where they do not work. They are being sold to hospitals, schools, police departments, and employers with promises that cannot be kept. And the people harmed by these failures are rarely the people buying the systems. The problem is not limited to bad actors. It is structural. Researchers publish papers with fundamental errors. Companies market products based on misleading claims. Journalists repeat press releases without scrutiny. The result is an ecosystem where hype feeds on itself, and skepticism is treated as obstruction. Consider predictive AI, which claims to forecast human behavior. These systems are used to predict which job applicants will succeed, which patients will develop complications, which students will drop out, and which defendants will reoffend. The underlying assumption is that past data can reveal future outcomes. But human lives are not governed by stable, discoverable patterns in the way these systems assume. The result is that predictive AI often…