A new preprint shows why AI-assisted psychiatric intake needs clinician-grounded quality assurance: better information recovery can coexist with unsupported inference and weak safety characterization.