Learning signatures in self-reported affect require introspection and are orthogonal to behavior

Do people learn to predict their feelings over time, and do such predictions manifest in behavior? Feeling ratings track with what we do. A better understanding of their properties may thus help elucidate behavior. Inspired by reinforcement learning frameworks, differences between expected and experienced feelings—affective prediction errors—have recently been added to the toolkit of behavioral prediction. But the extent of the analogy between affective prediction errors and conventional prediction errors about outcomes in the environment is unknown. Across a reanalysis of existing data (N = 4607) and three pre-registered experiments (N = 968; U.S. online samples collected in 2023), we probe affective prediction errors to document and dissect a core analogy: Learning reflected in decreasing (affective) prediction errors over time. We found that decreases in affective prediction errors depended on introspection, as prior experience with a task absent affective reports did not yield the same decreases (Experiment 1). A task manipulation forcing participants to alter their choices showed increased affective prediction errors, ruling out simple response alignment (i.e., to report feeling “as predicted”; Experiment 2). Decreases in affective prediction errors transferred across structurally similar tasks (i.e., stealing versus giving money; Experiment 3). Although affective prediction errors often tracked with behavior overall, their absolute decrease over time did not. In sum, we present evidence for convergence (i.e., learning and transfer) and divergence (i.e., introspection dependence and predictive epiphenomenality) between affective prediction errors and conventional prediction errors. Implications for the role of affective measures as a proxy for subjective value are discussed.