BACKGROUND There is substantial evidence that children with autism are impaired in face recognition. Although many researchers have suggested that this impairment derives from a failure of holistic face processing and a tendency to represent and encode faces on a part-by-part basis, this hypothesis has not been tested directly. METHOD Holistic face processing was assessed by comparing children's ability to recognize a face part (eyes, nose, or mouth) in the context of the whole face in which it was learned with their ability to recognize the same face part in isolation. RESULTS In Study 1, as expected, typically developing 9-year-olds (n = 27) and 11-year-olds (n = 30) were significantly better at recognizing face parts presented in the whole than in the part test condition, and this effect was limited to upright faces and not found for inverted faces. Consistent with prior findings, typically developing children were most accurate when face recognition depended on the eyes. In Study 2, high-functioning children with autism (n = 22) evidenced a whole-test advantage for mouths only, and were markedly deficient when face recognition depended on the eyes. Their pattern of performance diverged from age- and IQ-matched comparison participants (n = 20), who performed similarly to the typically developing children in Study 1. CONCLUSIONS These findings suggest that face recognition abnormalities in autism are not fully explained by an impairment of holistic face processing, and that there is an unusual significance accorded to the mouth region when children with autism process information from people's faces.