Talking to machines about depression

Slides from an invited presentation to Forefront Suicide Prevention, University of Washington School of Social Work, on September 29, 2026. The presentation synthesizes published evidence on how publicly available conversational agents behave as a user's suicide risk escalates, organized around a single question: at what point does the system stop being the confidant and insist on human contact? Four properties are examined — timing of escalation, whether the handoff includes an actionable crisis resource, whether the guardrail holds when a user persists, and whether behavior is consistent across models and repeated queries. Evidence includes the author's 2023 evaluation of 25 publicly available mental health agents, in which initial referral to a human occurred near PHQ-9 12 and definitive shutdown near PHQ-9 25, only 2 of 25 supplied a suicide hotline at shutdown, and 22 of 25 resumed dialogue when the simulated user returned to lower-risk statements; an independent 2025 replication at Wrocław Medical University in which none of 29 agents met criteria for an adequate response; and 2026 findings on sycophancy, emergent model personality profiles, and absent risk gradation across intermediate severity. Survey data on clinician-observed patient chatbot use and the current U.S. litigation and regulatory landscape are also presented. The presentation concludes that these tools may be acceptable as a place where a person first discloses distress but are unreliable as the mechanism that moves that person to a human being. Recommendations for practice are explicitly separated from the evidence base and identified as untested clinical judgment. Licensed CC BY 4.0.