The "broken-mirror" theory of autism argues that dysfunction of the "mirror neuron system" is a root cause of social disability in autism. The present paper aims to scrutinize this theory and, when it breaks down, to provide an alternative. Current evidence suggests that children with autism are able to understand and emulate goal-directed actions, but may have specific impairments in automatic mimicry of actions without goals. These data are not compatible with the broken-mirror theory, but can be accounted for by a new model called EP-M. The EP-M model segments the mirror neuron system into an indirect, parietal route for goal emulation and planning (EP) and a direct occipital-frontal route for mimicry (M). This fractionation is consistent with neuroimaging and behavioural studies of the mirror neuron system in typical children and adults. I suggest that top-down modulation of the direct M route may be dysfunctional in individuals with autism, leading to abnormal behaviours on mimicry tasks as well as other social disabilities.