JMIR Publications

ABSTRACT

Conversational mental health AI requires calibrated response and human escalation, not detection alone. Existing evaluations document missed crises, inconsistent handling of ambiguous disclosures, incorrect or delayed referrals, and abrupt safety messages that may suppress further disclosure. This Viewpoint introduces the Calibrated Response–Escalation Model, a layered, testable model that maps conversational evidence to proportionate response and human escalation while preserving autonomy and trust. Six revisable conversation-control states span figurative language to credible near-term danger; they guide response and routing rather than predict future behavior or determine access to care. Orthogonal modifiers distinguish self-directed risk, concern about another person, and exposure to suicide or self-harm content. The architecture separates safety-state inference, response policy, and escalation and routing, treats uncertainty as actionable, and pairs contextual escalation with repair and region-correct handoffs. These commitments operationalize SAMHSA's trauma-informed care principles of safety, trustworthiness, collaboration, and choice at the turn level. The contribution lies not in any single clinical concept but in integrating them into one auditable, turn-level response policy. The model accepts conservative high-acuity detection but shifts optimization to the response layer, making false positives less harmful rather than assuming they are harmless. The model is a conceptual specification, not a demonstrated safety solution; its value depends on independent validation, clinician and lived-experience co-design, and governance capable of monitoring real-world adverse events.