Human-robot collaboration systems benefit from the ability of the robot to recognize people’s intentions. People’s nonverbal behavior while performing tasks, especially gaze, has shown to be a reliable signal to recognize what people intend to do. We propose an additional usage of this signal: to recognize when something unexpected has occurred during the task. Case studies from a dataset of gaze behavior when controlling a robot indicate that people’s gaze deviates from ordinary patterns when unexpected conditions occur. By using such a system, robot collaborators can identify unexpected behaviors and smoothly take corrective action.