The Stable-World Principle

The stable-world principle is Gerd Gigerenzer's compression of a lifetime's research into a single decision procedure: before deploying any algorithm in any domain, ask whether the world in which the algorithm will operate is stable or unstable. A stable world is one where the future resembles the past closely enough that patterns learned from history continue to hold—the rules of chess, the geometry of faces in photographs, the grammar of a natural language. In stable worlds, more data and more computation reliably yield better performance, and the machines are genuinely and durably superior. An unstable world is one where the relationships that held yesterday may not hold tomorrow, because the system includes agents who adapt, markets that respond, and feedback loops that alter the very regularities being measured—predicting human behavior, epidemic trajectories, financial markets, political outcomes. In unstable worlds, Gigerenzer has shown, simple rules using little data often beat complex algorithms using much, because the complex algorithm's parameter estimates are drawn from