Defining mental health use of AI: what counts?

Clinical takeaway: Most mental health use of chatbots looks like everyday emotional conversation rather than explicit help-seeking, so a patient who says they "talk to AI" may mean anything from venting to crisis disclosure; establishing which is the first step. 

Chatbot use for emotional support is now common enough that clinicians will encounter it, but analyses of how common, or what to do next, remain limited. Some suggest mental health moments in AI conversations are vanishingly rare; others argue emotional reliance on chatbots is near-routine. 

OpenAI has concluded that 0.01% of messages in a given week showed possible signs of mental health emergencies involving psychosis or mania, while a Common Sense Media survey reported that 72% of teenagers have used AI chatbots as companions. Whether those figures are competing measurements of one behavior or answers to entirely different questions has been unclear. The distinction matters, because duties of care and safety evaluation thresholds hinge on whether an interaction counts as mental health. A cross-sectional analysis of a large public corpus of chatbot conversations tested that directly. 

A clear answer depends almost entirely on where the line is drawn, concluded an assessment of 620,699 publicly available ChatGPT conversations. Under a conservative definition, requiring an identifiable real person seeking help for a psychological problem, 0.21% of conversations were mental health related. Under an expansive definition, requiring only that mental health content be present regardless of who it concerned or whether help was sought, the figure was 4.90%, a 23-fold difference from the same conversations scored by the same classifier. 

The two definitions also captured different kinds of conversations. Those meeting the conservative bar scored higher on topicality and help-seeking intent, consistent with its requirement for explicit help-seeking. Conversations captured only by the expansive definition tended to touch mental health topics without centering an identifiable person or a request for guidance, closer to emotional and interpersonal territory than to anything resembling a clinical encounter. 

The conversations came from WildChat-4.8M, a public corpus of human-chatbot exchanges, scored by a large language model classifier on four domains: centrality of mental health content, help-seeking, clinical terminology, and severity of distress. Against 370 human-coded conversations, the classifier showed 87% sensitivity and 95% specificity, though a positive predictive value of 67% means the prevalence figures run softer than the definitional gap itself. The corpus captures only conversations users chose to make public. 

The authors' proposed fix is a tiered taxonomy for future surveillance of AI conversations, one that separates crisis or high-risk exchanges, clinically framed help-seeking, and broad affective or interpersonal use. Each tier carries its own clear need: crisis conversations need reliable detection and escalation safeguards, clinically framed ones need evaluation for accuracy and referral behavior, and broad emotional use needs attention to dependence and developmental appropriateness. 

"Standardized reporting would help researchers, clinicians, and regulators benchmark chatbot performance, monitor changes as models and platform policies evolve, identify safety gaps, and determine where oversight is warranted," the authors conclude. 

Source: McBain RK, et al. (2026 Aug 24) JAMA Netw Open. Prevalence of Mental Health Discussions in Publicly Available Generative AI Conversations