Abstract Previous eye tracking studies have reported that children with autism spectrum disorders (ASD) fixate less on faces in comparison to controls. To properly understand social interactions, however, children must gaze not only at faces, but also at actions, gestures, body movements, contextual details, and objects, thereby creating specific gaze patterns when observing specific interactions. We presented three different movies of social interactions to 111 children (71 with ASD) who watched each of the movies twice. Typically developing children viewed the movies in a remarkably predictable and reproducible manner, exhibiting gaze patterns that were similar to the mean gaze pattern of other controls, with strong correlations across individuals (inter-subject correlations) and across movie presentations (intra-subject correlations). In contrast, children with ASD exhibited significantly more variable/idiosyncratic gaze patterns that differed from the mean gaze pattern of controls and were weakly correlated across individuals and presentations. Most importantly, quantification of gaze idiosyncrasy in individual children, enabled separation of ASD and control children with higher sensitivity and specificity than traditional measures such as time gazing at faces. Individual magnitudes of gaze idiosyncrasy were also significantly correlated with ASD severity and significantly correlated across movies and movie presentations, demonstrating their clinical sensitivity and reliability. These results suggest that gaze idiosyncrasy is a potent behavioral abnormality that characterizes many children with ASD and may contribute to their impaired social development. Quantification of gaze idiosyncrasy in individual children may aid in assessing their ASD severity over time and in response to treatments. Lay Summary Typically developing children watch movies of social interactions in a reliable and predictable manner, attending faces, gestures, body movements, and objects that are relevant to the social interaction and its narrative. Here, we demonstrate that children with ASD watch such movies with significantly more variable/idiosyncratic gaze patterns that differ across individuals and across movie presentations. We demonstrate that quantifying this variability is a very potent way of identifying children with ASD and determining the severity of their social ASD symptoms.