Violating statistical structure impairs detection of deviant and incidental events.

Learning about the statistical structure of our environment is thought to shape perception, but it is unclear how. A recent theory suggests that percepts are initially biased toward the expected, with particularly unexpected observations triggering reactive sensory gain increases-balancing requirements for fast and accurate perception alongside reliable sensory estimates for model updating. Six experiments tested this account, where participants detected visual stimuli on the circumference of, and at the center of, a circle. Circumference stimuli followed a spatial or orientation regularity, which then changed abruptly. Bayesian changepoint modeling showed that hit rates were lower for all events following such disruption of the learned probabilistic structure (hereafter "surprise"). Performance recovery after one change took several trials but became immediate when changes were more frequent. These findings suggest broad perceptual facilitation of the expected regardless of latency, and we thus consider how models may be accurately updated when the world changes, despite poorer perception.