Confidence exhibits systematic individual differences across mental health, gender, and age. However, it remains unknown whether these distinct sources of metacognitive bias have common or distinct computational origins.

To address this question, we developed a novel dynamic computational model of metacognition to study the temporal evolution of underconfidence associated with individual differences in transdiagnostic anxiety symptoms and gender in samples of online participants (total N = 1,447).

We found that underconfidence associated with anxiety symptoms became more prominent the longer individuals took to make metacognitive judgments – suggesting that it is exacerbated by additional time for introspection. In contrast, gender-related underconfidence decreased with greater metacognitive judgment time – suggesting that additional time for introspection is able to remediate prepotent biases. Our computational model of confidence explained these effects – while both gender and anxiety symptoms involved shifts in confidence criteria, only anxiety symptoms involved a temporal accumulation of negatively biased evidence about one’s ability.

Our study reveals multiple computational pathways to the formation of underconfidence, in turn highlighting specific potential mechanisms for its remediation.