AI safety seems hard to measure

Abstract

Four analogies for why “We don’t see any misbehavior by this AI” isn’t enough.