Abstract
Appropriate AI reliance requires users to accurately evaluate their own performance in order to discern whether to retain or defer responsibility. While underconfidence is a known driver of automation bias, little is known about how collaboration with AI itself influences users’ metacognition. We investigated how cognitive offloading different stages of the decision-making process affects confidence calibration. Participants completed diagnostic decision-making tasks with varying levels of memory and judgement support. Overall, cognitive offloading impaired confidence calibration, with unaided decision makers showing the greatest alignment between confidence and accuracy. Importantly, offloading different cognitive processes produced distinct metacognitive biases: judgement offloading led to overconfidence, whereas combined offloading of memory and judgement processes led to underconfidence. These findings demonstrate that AI support can disrupt user confidence calibration in systematic ways, depending on the type and extent of cognitive delegation. The results highlight metacognitive miscalibration as a critical and underexplored consequence of human-AI collaboration.