Ambiguity and confirmation bias in reward learning

We tend to interpret feedback in ways that confirm our pre-existing beliefs. Such confirmation bias is often viewed as a cognitive flaw, but may have adaptive foundations. We propose both a novel experimental paradigm and Bayesian computational model to examine how confirmation bias shapes reward learning. When outcomes are ambiguous, forming accurate assessments requires drawing on all available information, including prior beliefs. Confirmation bias may thus constitute an inductive bias that enhances learning, analogous to missing data imputation. We develop and test this theory using a reward learning task in which information about outcome valence (but not magnitude) is withheld, allowing greater scope for subjective interpretation. Our Bayesian model explains the dynamics of behavior and stated beliefs better than a more traditional learning model. Moreover, stated beliefs about the positivity of ambiguous outcomes are correlated with optimism. Together, these findings suggest an adaptive role for confirmation bias, and link individual interpretations of ambiguity to broader dispositional tendencies.