Conference on Cognitive Computational Neuroscience

A pressing question concerns how individuals learn from potentially misleading feedback, which is a common occurrence in social settings. We addressed this question using an online modified version of the two-armed bandit task, that informed participants (n=107) about reward-outcomes of their choices through “feedback agents” which varied in how frequently they lied (i.e., in their reliability). Computational modelling of participants’ behavior revealed that credit assignment (i.e., the extent of value-update following reward feedback) increased as a function of feedback-reliability. Strikingly, while credit assignment for the always-truthful agent was symmetric (i.e., equal for positive and negative outcomes), unreliable feedback elicited a positivity bias such that credit assignment was greater for positive compared to negative feedback. Critically, this bias could not be accounted for by optimal Bayesian learning models. Finally, exploratory analyses suggest that obsessive-compulsive symptoms, paranoid thoughts and dogmatism are related to distinct biases in learning from disinformation. Our results can help explain how learning from disinformation might contribute to deficits in mental health and social communication (e.g., political radicalization and belief in conspiracy narratives).