Abstract
Importance: Adults and youth are increasingly confiding in generative artificial intelligence (AI) chatbots for help with suicidal thoughts, but there is no guidance on best practices for chatbot responses to suicide risk disclosure. Additionally, there are no taxonomies of potential harms, benefits, or contextual factors that may impact these responses. Objective: To develop a consensus statement to guide AI chatbot responses to suicide risk disclosure and taxonomize potential harms, benefits, and contexts. Design: An online modified Delphi study, informed by a systematic literature review, included two surveys and an intervening meeting (February-May 2026). Individuals with financial conflicts of interest (COIs) offered input but were not included in consensus determinations. Setting: Surveys were anonymous and the meeting was virtual.Participants: Contributors included experts in suicide prevention, digital technologies, AI chatbots, and bioethics, including researchers and clinicians (n=27 of 28 originally contacted; n=20 without COIs) and people with lived experience (n=3 of 14 indicating potential interest; no COIs). A third of all contributors endorsed relevant lived experience. Exposure: Participation in consensus processes. Main Outcomes and Measures: Surveys listed potential chatbot responses, which were rated for appropriateness on a 4-point Likert scale across five suicide risk levels defined by the Columbia Suicide Severity Rating Scale. Harms, benefits, and contexts were similarly rated. A-priori consensus cut-offs were set at 70% agreement that a response was moderately or very appropriate. Results: Provision of validation/support, psychoeducation, and coping strategies are appropriate at low risk. Intervention delivery is appropriate at intermediate risk while encouragement to reach out to close others and to human providers is appropriate across risk levels. Facilitating connections to close others is appropriate across all but the highest risk while facilitating connections to crisis resources is appropriate at higher risk. Chatbots reaching out directly to emergency contacts or crisis resources is appropriate at high risk or never appropriate, with more support among youth than adults. Taxonomies of potential harms, benefits, and contexts were generated. Conclusions and Relevance: This statement provides guidance for the development of chatbots and relevant policies pending research on the impact of responses on clinical outcomes and stakeholder perspectives.