assistance game
An assistance game is an AI safety framework in which an AI tries to satisfy human preferences while remaining uncertain about what they are.
An assistance game is an AI design framework proposed by Stuart Russell and his collaborators. It models a human and an AI as two players in one game: the AI's only goal is to satisfy the human's preferences, but it does not know for certain what those preferences are and infers them from human behavior.
Unlike designs that maximize a fixed objective, an AI that is uncertain about preferences has an incentive to ask for clarification and to accept being stopped by a human. Research on the off-switch game, which studies whether humans can switch an AI off, belongs to this framework.
This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.