Moment-to-moment neural variability plays a crucial role in learning and behavioral flexibility. This variability has been linked to age-associated differences in task performances and resting-state activity. However, its directional effects, whether increased or decreased neural variability is beneficial, remain context-dependent and unclear. Here, we demonstrate that changes to cortical neural variability during naturalistic experience serve as a readout of changes to region-specific computations, distinct from resting-state variability and mean neural activity. During the affective experience, older adults exhibited a bidirectional neural variability shift within the medial and lateral orbitofrontal cortex. By employing a Bayesian learning model, we revealed that this corresponded to heightened uncertainty in the computational estimates of evolving valence in older adults. This was reflected in frequent state switching in neural response and temporally distorted representations within the lOFC in older adults, compared to stable, coherent representations in the young. Crucially, these representational changes mirrored the arousal responses of the two groups. These results advance our understanding of how neural variability might carry unique information in aging, particularly to their changing and possibly idiosyncratic affective experiences.