Does the human visual system contain a specialized system for face recognition, not used for the recognition of other objects? This question was addressed using the "face inversion effect" which refers to the loss of our normal proficiency at face perception when faces are inverted. We found that a prosopagnosic subject paradoxically performed better at matching inverted faces than upright faces, the opposite of the normal "face inversion effect". The fact that his impairment was most pronounced with the stimuli for which normal subjects show the greatest proficiency in face perception provides evidence of a neurologically localized module for upright face recognition in humans. An additional implication of these data is that specialized systems may control behavior even when they are malfunctioning and therefore maladeaptive, consistent with the mandatory operation of such systems according to the "modularity" hypothesis of the cognitive architecture.