Proceedings of the Annual Meeting of the Cognitive Science Society

A pivotal advantage of human intelligence lies in metacognitive monitoring. This ability allows individuals to detect uncertainty in their judgments and adapt their behavior accordingly, such as deferring decision-making or seeking help. However, the computational mechanism underpinning metacognitive processing still remains incompletely characterized. We propose a risk-avoidance-based computational theory of metacognition, stating that second-order confidence is formed through an asymmetric strategy: individuals conservatively downweight potential gains while overestimating potential risks inherent in their decision outputs. We formalize this principle into a dual channel conservative confidence (DCCC) framework and instantiate it within a large language model (LLM). Evaluations across high-risk medical and legal judgment tasks demonstrate that the model's confidence signals closely align with human metacognitive profiles, including patterns in the distribution of confidence levels, patterns of error detection, and help-seeking propensities. These findings not only delineate a tractable pathway for implementing machine metacognition but also reveal a core risk-avoidance-oriented computational principle that governs human metacognition processing.