Disinformation is often considered to pose a threat to open societies. However, we know little regarding the learning biases elicited by disinformation. To address this, we developed a novel reinforcement learning task wherein participants chose between lotteries without knowing the true outcomes of their choices (rewards or non-rewards). Instead, they received choice-feedback from sources who occasionally disseminated disinformation by lying about choice outcomes. As these sources varied in their truthfulness this allowed us to test how learning differed based on source-credibility. Across two experiments computational modelling indicated that learning increased in tandem with source-credibility, consistent with normative Bayesian principles. However, we also observed striking biases reflecting divergence from normative learning patterns. Notably, individuals learned from sources known to be unreliable and increased their learning from trustworthy information when it was preceded by non-credible information. Furthermore, the presence of disinformation exacerbated a “positivity bias” whereby individuals self-servingly boosted their learning from positive, compared to negative, choice-feedback. Our findings reveal cognitive mechanisms underlying learning biases in the face of disinformation, with potential implications for strategies aimed at mitigating its pernicious effects.