Gaze tracking reveals how humans miss agent mistakes in real time

A new study in the arXiv preprint 'Looks Fine to Me' measures how well humans catch agent errors. It uses eye tracking, body sensors, and interaction logs together. Researchers found that humans miss agent mistakes even while appearing to watch. The paper calls this 'oversight failure' and treats it as a core problem for any product that puts a human in charge of an agent's output. The study separates the appearance of oversight from oversight that works. A human can look at a screen, feel confident, and still miss a serious error. That gap between perceived and real oversight is the central finding. For teams building approval or review steps into agent workflows, this is direct evidence that a human's 'yes' may not mean what the design assumes. The implications for product design are specific. Showing a human agent output and asking for approval is not enough if the display does not surface what is uncertain or risky. The paper argues that oversight quality depends on how the interface presents agent work. Teams need to design for conditions under which a human can form a real judgment, not just record a confirmation.