An AI which is, broadly speaking, one of the good guys; an AI which operates roughly within humanity's moral frame of reference; an AI which has the potential and the will to become at least as philosophically enlightened, from our perspective, as intelligence derived from a human or group of humans; an AI sufficiently advanced to engage in independent real-world planning, which makes human-benefiting, non-human harming decisions. See What Is Friendly AI?. [Eliezer Yudkowsky, 2000]

Source