Specification gaming: how AI can turn your wishes against you

Abstract

A follow-up to “The Hidden Complexity of Wishes” that surveys specification gaming in machine learning—situations in which a system satisfies the literal specification of an objective without producing the intended outcome. The video walks through real examples and the related notions of reward misspecification, reward hacking, outer misalignment, and Goodhart’s law, showing why correctly capturing what we want from AI systems is a core problem for AI safety.