Metacognition allows us to explicitly represent the uncertainty in our perceptions and decisions. Recent theories suggest that we use predictive models of our environment to optimise these introspective processes, but extant accounts disagree about the role prediction plays - some accounts suggest that we should have more sensitive subjective insight for predictable events, while others stress that metacognition should be enhanced for surprising prediction errors. Here two experiments compare these accounts. Participants performed actions to generate visual outcomes that could move in expected or unexpected directions. Across both experiments, signal detection analyses revealed enhanced metacognition for unexpected outcomes. A combination of reverse correlation and computational modelling suggested this advantage arose because metacognitive processes are more sensitive to unexpected information. These results are consistent with higher order inference models of introspective awareness and point to a mechanism that may optimise diverse aspects of cognition and behaviour in an unstable world.