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

Arvind Narayanan and Sayash Kapoor expose a deliberate confusion: the term "artificial intelligence" now covers both ChatGPT's genuine text generation and COMPAS's courtroom predictions that perform no better than a two-variable rule on an index card. The Fragile Families Challenge proved the point—160 research teams with fifteen years of data couldn't predict children's outcomes better than coin flips. A supportive neighbor feeding a struggling child blueberries changed everything; no algorithm captured her. From the Dutch welfare scandal that toppled a prime minister to Mount St. Mary's president ordering faculty to "drown the bunnies," this book names the fraud.

You'll learn:

- Why 137 variables produce results a four-feature model matches

- How cumulative advantage makes prediction theoretically impossible

- What the FTC means when it calls predictive AI claims "deceptive"