The Direct Approach of Testing for AGI-Consciousness

Consciousness and embodied intelligence are two major stumbling blocks on the way from AI to Artificial General Intelligence (AGI). While testing for embodied intelligence (e.g. the ability of a robot to do diverse household chores) is obvious, testing an AI agent for consciousness is beyond the current state of the art. The problem is compounded by the fact that AI agents are known to behave deceptively, and therefore querying the agent about its consciousness is unreliable. This paper introduces a mechanism, Consciousness Notification (CN), which detects the emergence of consciousness in an AI agent; upon detection, CN informs the agent’s owner, the Authority. CN is inspired by the connection in humans between emotions and physiology. In contrast with existing approximate, similarity-based methods, CN is a novel and direct approach in the sense that CN is embedded in the AI agent. The paper also introduces requirements that are necessary for a direct mechanism to be sound, and a theory by which it formally proves that, under certain conditions, the CN mechanism satisfies these requirements. The conditions formally capture the type of cheating that the AI will have to perform to evade the CN mechanism.