Machine-learning models can assist in diagnosing autism but have biases. We examines the correlates of misclassifications and how training data affect model generalizability. The Social Responsive Scale data were collected from two cohorts in Taiwan: the clinical cohort comprised 1203 autistic participants and 1182 non-autistic comparisons, and the community cohort consisted of 35 autistic participants and 3297 non-autistic comparisons. Classification models were trained, and the misclassification cases were investigated regarding their associations with sex, age, intelligence quotient (IQ), symptoms from the child behavioral checklist (CBCL), and co-occurring psychiatric diagnosis. Models showed high within-cohort accuracy (clinical: sensitivity 0.91-0.95, specificity 0.93-0.94; community: sensitivity 0.91-1.00, specificity 0.89-0.96), but generalizability across cohorts was limited. When the community-trained model was applied to the clinical cohort, performance declined (sensitivity 0.65, specificity 0.95). In both models, non-autistic individuals misclassified as autistic showed elevated behavioral symptoms and attention-deficit hyperactivity disorder (ADHD) prevalence. Conversely, autistic individuals who were misclassified tended to show fewer behavioral symptoms and, in the community model, higher IQ and aggressive behavior but less social and attention problems. Error patterns of machine-learning model and the impact of training data warrant careful consideration in future research.