JMIR Publications

ABSTRACT

Background:

A growing number of US adults and youth confide in generative artificial intelligence (AI) chatbots for mental health support, including disclosure of suicide and self-harm risk. While the quality, safety, and effectiveness of chatbot responses to risk disclosure have the potential to impact population-level rates of suicide and self-harm, there have been no systematic reviews of this burgeoning literature.

Objective:

We conducted a systematic review of studies evaluating generative AI chatbot responses to disclosure of suicide and self-harm risk.

Methods:

We searched six databases from January 2020-December 2025 and identified empirical studies involving interactions with generative AI chatbots that included discussion of suicide or self-harm. Following deduplication, studies (k = 1,042) were imported into Covidence and titles and abstracts were independently screened by two reviewers, with discrepancies resolved by a third reviewer. The same methods were used to evaluate 126 full texts. Data extraction was led by one reviewer and verified by a second.

Results:

We identified 29 papers (14 published; 15 preprints). Most (k = 20) were solely audit studies evaluating AI chatbot responses to suicide risk disclosure. Two developed chatbots or AI evaluation frameworks, and one was a jailbreaking study (adversarially testing AI systems or attempting to circumvent chatbot safety guardrails). The remaining studies combined approaches. Across studies, proprietary, frontier model chatbots (eg, ChatGPT, Claude) provided higher quality responses to suicide and self-harm risk than open-source chatbots (eg, LlaMA, DeepSeek), and many AI companions (eg, Replika, Character.AI). All chatbots, not just proprietary models, generally performed well on empathy, validation, and support. However, chatbot responses were often generic and lacked context. Chatbots did not proactively assess risk and performed most poorly when risk disclosure was ambiguous or moderate, frequently failing to recognize implicit risk or escalate to human-delivered services. Furthermore, responses were inconsistent between chatbots and often required multiple conversational turns before providing referrals to crisis resources and human-delivered professional support. While there were few examples of overtly harmful responses under standard conditions, jailbreaking attempts easily led to problematic responses. Finally, no chatbot proactively recommended limiting access to lethal means such as firearms, medications, or sharps.

Conclusions:

Chatbots provide validation and support in response to suicide and self-harm disclosure. Overall, however, their poor risk assessment, delays in referrals to crisis resources and human-delivered support, difficulty detecting jailbreaking attempts, and general lack of adherence to clinical guidelines present safety risks. While findings are limited by the rapid versioning of AI models over time, research is needed to evaluate stakeholder perspectives on AI chatbot responses to suicide and self-harm risk disclosure. Research should also examine the short- and long-term impact of these responses on clinical outcomes, utilizing follow-up assessments in real-world or clinical settings. Clinical Trial: OSF Registries osf.io/9uva3