Textbook of AI

The frame problem, articulated by John McCarthy and Pat Hayes in their 1969 paper Some Philosophical Problems from the Standpoint of Artificial Intelligence, is the problem of specifying, in logic, the unchanging consequences of an action without listing them all explicitly. If an agent moves a block from one location to another, the action changes the block's location, but logically, one must also specify that the colour of the block does not change, that other blocks remain where they were, that the agent is still in the same room, that the temperature is unaffected, and so on, for every possible property and every other entity in the world.

The technical frame problem is one of representation: how to encode action effects compactly so that "default persistence" of all unmentioned facts is implicit. Frame axioms that explicitly list non-changes blow up the knowledge base intolerably. Various solutions have been proposed: STRIPS-style add/delete lists with the implicit "STRIPS assumption", McCarthy's circumscription, Reiter's default logic, Sandewall's chronological minimisation, and Reiter's later successor-state axioms.

The philosophical frame problem, as discussed by Daniel Dennett (1984) and Jerry Fodor (1987), is broader: how does any cognitive agent decide which of its enormous body of background knowledge is relevant to a given task without first checking all of it? This version of the frame problem is unsolved in classical symbolic AI; modern large language models partially sidestep it by learning, implicitly, which knowledge is contextually relevant.