Social Prediction Error Signals During the Trust Game in Patients with Schizophrenia: An fMRI Study

Schizophrenia (SZ) is characterized by marked deficits in social cognition, yet the computational and neural mechanisms underlying social reinforcement learning remain poorly understood. We investigated dynamic social learning in patients with SZ (N = 29) and healthy controls (N = 45) during an iterative Trust Game performed under functional magnetic resonance imaging (fMRI).

Participant behavior was modeled using a Rescorla-Wagner framework to estimate learning rate ) and decision precision (inverse temperature, β), alongside trial-by-trial prediction errors (PEs) for neural analyses.

Behaviorally, patients exhibited a marked impairment in utilizing positive social feedback, reflected in reduced win-stay behavior, significantly lower learning rates, and diminished model fit. Critically, patient behavior was best explained by a null model, indicating a failure to consistently integrate feedback into decision-making. At the computational level, reduced decision precision (lower β) was significantly associated with positive symptom severity, suggesting a link between stochastic choice behavior and psychosis. Neuronally, while healthy controls showed robust PE-related activation in canonical reward and visual processing regions, patients exhibited significantly attenuated PE signaling, particularly within the left lateral occipital cortex and fusiform gyrus.

These findings indicate a fundamental disruption in precision-weighted social belief updating in schizophrenia, characterized by stochastic decision-making and degraded neural encoding of prediction errors. This supports predictive coding accounts of psychosis, in which impaired precision assignment to sensory feedback undermines adaptive learning and contributes to positive symptoms.