Suspicious Minds: How Confirmation Bias Shapes Learning And Belief Updating In The Age Of Misinformation

Introduction: Belief persistence in the face of contradictory evidence is a hallmark of human cognition, particularly in online environments where misinformation proliferates. While confirmation bias has been widely implicated in this phenomenon, its interaction with Reinforcement Learning (RL) dynamics and individual cognitive styles remains poorly understood. Methods: We examined how reward structures, prior confidence, and psychological traits shape epistemic judgments and belief updating. We designed a three-phase experiment combining fake news classification, a probabilistic learning task with varying rewards, and a belief revision phase. Real and fake headlines were judged for veracity and confidence, then used as stimuli in a two-armed bandit task. Reinforcement was probabilistically tied to either perceived truthfulness or confidence. We compared learning models and collected psychometric data to link learning strategies to individual traits. Results: By modelling belief persistence as the outcome of multiple interacting systems within a single, ecologically grounded framework we show how participants' behaviour was shaped by the alignment between reward and prior beliefs. When rewards confirmed prior judgments, accuracy increased; when rewards conflicted with expectations, learning declined. Computational modelling revealed stable inter-individual preferences for generalization or independent learning, modulated by intuitive thinking, conspiracist beliefs, and metacognitive confidence. Belief revision was rare and mostly confined to cases of low initial confidence. Conclusions: These findings reveal how belief stability and change emerge from the interplay of reinforcement structure, cognitive dispositions, and confidence. By unifying belief dynamics with RL frameworks, our study offers a formal account of misinformation resilience and lays the groundwork for targeted interventions in digital environments.