JMIR Publications

ABSTRACT

Artificial intelligence (AI) is entering mental health care along two paths. One is visible: systems supporting screening, formulation, treatment selection, and between-session monitoring. The other is quieter: consumer tools used for reassurance, advice, or companionship before a clinician is involved. These paths meet in the patient's life and warrant the same standards. Mental health care is a demanding test for AI because outcomes are contextual, assessment depends on what a person can report, the therapeutic relationship is itself part of treatment, and harm may accumulate even when no single output appears unsafe. We propose two standards that apply to every system. Fitness for use asks whether AI converts a patient signal into timely, equitable, and accountable clinical action: whether performance survives external validation, remains calibrated across subgroups, provides net benefit, and connects to a responsive pathway. Longitudinal safety asks whether that judgment survives months of use, as models and populations drift, clinicians adapt to repeated outputs, and relationships with conversational systems accumulate. A prediction model may fail because no clinical pathway turns its output into care; a therapeutic chatbot may fail because repeated replies erode help-seeking or substitute for alliance. Either system can fail either standard. Drawing on evidence across risk stratification, digital therapeutics, between-session monitoring, and consumer conversational AI, we connect reporting guidance, implementation science, and longitudinal evaluation within a framework organized around clinical decisions. The question is not how a system enters care, but whether its effects remain beneficial, equitable, and accountable across the trajectory of use.