# Imitation-based agent

An imitation-based advanced agent (or pseudoagent) is an AI whose central loop, rather than observe-predict-act as in a standard agent, is “output the behavior that matches as closely as possible what a reference human would do”. The central idea is that, (a), this is a much easier behavior to define in terms of today’s supervised learning paradigms, and (b), maybe this gets rid of many undesirable behaviors that (b1) stem from being a consequentialist agent with goals and (b2) are clearly not what the reference human would do. An archetypal kind of act-based agent. Invented by Paul Christiano as the base case of larger proposals to have AIs recursively imitate human+AI systems.