Social cognition has become a major focus in psychosis research aimed at explaining heterogeneity in functional outcome and developing interventions oriented to functional recovery. However, there is still no consensus on the structure of social cognition in psychosis, and research in this area has been plagued by lack of replication. Our first goal was to replicate the factor structure of social cognition using nearly identical tasks in independent samples. Our second goal was to externally validate the factors as they relate to nonsocial cognition and various symptoms in the prediction of functioning using machine learning. Confirmatory factor analyses validated a three-factor model for social cognition in psychosis (low-level, high-level, attributional bias factor). A least absolute shrinkage and selection operator regression and cross-validation provided evidence for external validity of data-driven linear models including the social-cognitive factors, nonsocial cognition, and symptoms. We addressed the replicability problems that have impeded research in this area, and our results will guide future psychosis studies.