This paper develops and estimates a closed-form model of Bayesian learning of subjective survival beliefs within the framework of Choquet decision theory. Data from the Health and Retirement Study (HRS) indicate that, on average, young respondents underestimate their true survival probability whereas old respondents overestimate their survival probability. Such subjective beliefs violate the rational expectations paradigm and are also not in line with the predictions of the rational Bayesian learning paradigm which implies convergence of subjective to underlying ‘objective’ probabilities. Based on the assumption of non-additive beliefs, we therefore introduce a model of Bayesian learning which combines rational learning with the possibility that the interpretation of new information is prone to psychological attitudes. We estimate the parameters of our theoretical model by pooling the HRS data. Despite a parsimonious parametrization we find that our Choquet model results in a remarkable fit to the average subjective beliefs expressed in the data.