AI should be redefined to prioritize human objectives, judgment and oversight, departing from the standard optimization model, in order to manage the existential risks posed by the pursuit of superhuman AI.