AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference

Arvind Narayanan and Sayash Kapoor categorize artificial intelligence into three distinct domains to explain why the technology frequently fails to meet its promises. They distinguish between generative AI, which manipulates digital content, and predictive AI, which attempts to forecast human behavior in fields like hiring, policing, and healthcare. The authors argue that while large language models show genuine capabilities, predictive algorithms often function as digital astrology because they rely on historical data that cannot account for the complexity of future social outcomes. By dismantling the technical fallacies behind automated decision-making, the book exposes how companies use the label of "intelligence" to mask statistical limitations and flawed logic.

Data scientists, policymakers, and corporate leaders read this book to develop a technical skepticism toward commercial software claims. Readers learn to identify common pitfalls such as data leakage and the confusion between correlation and causation. The text provides a practical framework for questioning algorithm developers about accuracy and bias. Anyone tasked with implementing or regulating new technology will exit with the tools to separate functional automation from speculative marketing, ensuring they do not deploy software that produces harmful or inaccurate social interventions.