On average, “young” people underestimate whereas “old” people overestimate their chances to survive into the future. Such subjective survival beliefs violate the rational expectations paradigm and are also not in line with models of rational Bayesian learning. In order to explain these empirical patterns in a parsimonious manner, we assume that self-reported beliefs express likelihood insensitivity and can, therefore, be modeled as non-additive beliefs. In a next step we introduce a closed form model of Bayesian learning for non-additive beliefs which combines rational learning with psychological attitudes in the interpretation of information. Our model gives a remarkable fit to average subjective survival beliefs reported in the Health and Retirement Study.