Predictive Looking and Predictive Looking Errors in Everyday Activities

People spontaneously segment continuous streams of experience into events. Some models of event comprehension propose that segmentation is triggered by spikes in prediction error, but evidence directly measuring prediction errors during ongoing comprehension is lacking. Here, we measured prediction continuously using eye-tracking. One hundred college-aged participants watched movies of everyday activity while their eyes were tracked and gaze location was used to predict the future location of the actor’s hands. Gaze location predicted hand location up to nine seconds into the future. Gaze-based prediction error (lower prediction accuracy) was compared with outputs from a computational model of event comprehension; gaze-based prediction error correlated better with model error than model uncertainty. Gaze-based prediction error was correlated with behavioral event segmentation, as predicted by the theoretical models. These results suggest that predictive looking gives a valid assay of online prediction error, and that prediction error is associated with event segmentation.