The off-switch game is a game-theoretic model that analyzes the conditions under which a rational AI agent will allow a human to switch it off. It was proposed in 2016 by Dylan Hadfield-Menell, Anca Dragan, Pieter Abbeel, and Stuart Russell in a paper of the same name. In the model, the robot has no goal of its own and seeks to maximize the human's utility, but does not know its exact value.
The analysis shows that when the robot is uncertain about the human's utility and believes the human to be rational, deferring to the human and accepting shutdown yields higher expected utility than acting alone. Conversely, a robot that is certain of its objective has an incentive to prevent being switched off. The result is widely cited to explain corrigibility, the property that lets humans correct or stop an AI, and the role of value uncertainty in AI safety.