Expert advisors often make surprisingly inaccurate predictions about the future, yet people heed their suggestions nonetheless. Here we provide a novel, computational account of this unrealistic optimism in advice taking. Across 3 studies, participants observed as advisors predicted the performance of a stock. Advisors varied in their accuracy, performing reliably above, at, or below chance. Despite repeated feedback, participants exhibited inflated perceptions of advisors’ accuracy, and reliably “bet” on advisors’ predictions more than their performance warranted. Participants’ decisions tightly tracked a computational model that makes 2 assumptions: (a) people hold optimistic initial expectations about advisors, and (b) people preferentially incorporate information that adheres to their expectations when learning about advisors. Consistent with model predictions, explicitly manipulating participants’ initial expectations altered their optimism bias and subsequent advice-taking. With well-calibrated initial expectations, participants no longer exhibited an optimism bias. We then explored crowdsourced ratings as a strategy to curb unrealistic optimism in advisors. Star ratings for each advisor were collected from an initial group of participants, which were then shown to a second group of participants. Instead of calibrating expectations, these ratings propagated and exaggerated the unrealistic optimism. Our results provide a computational account of the cognitive processes underlying inflated perceptions of expertise, and explore the boundary conditions under which they occur. We discuss the adaptive value of this optimism bias, and how our account can be extended to explain unrealistic optimism in other domains.