Thseen Nazir
Large language models (LLMs) such as ChatGPT are increasingly used around mental health, yet design-oriented syntheses remain limited. We conducted a PRISMA-aligned systematic scoping review focused specifically on ChatGPT (GPT-3.5/4+) in mental health-related use including "in the wild" adoption by laypeople, training applications, and clinical-adjacent pilots. Searches of five databases (November 2022 to August 2025), with citation tracking and gray-literature screening, yielded 34 studies spanning randomized and non-randomized experiments, pilot trials, surveys/interviews, simulations, digital ethnography, and structured editorials. Evidence supports adjunct, not replacement, roles. In education/supervision, one randomized trial and a comparative supervision study show skill gains when practice is scaffolded with rubrics and human oversight; expert ratings judged trainee case conceptualizations acceptable. In clinical/adjacent contexts, signals include quality-of-life improvement (small inpatient pilot), short-term anxiety reduction when the model provides empathetic feedback, and a clinical RCT (outside psychiatry) showing reduced anxiety/depression with a ChatGPT adjunct. Studies of public/self-help use document appropriation of ChatGPT as a "digital therapist" with identified risks including privacy concerns, boundary violations, and over-reliance. Safety-critical tasks remain unreliable (e.g., under-identification of suicide risk, degradation with complexity, and cultural-fit gaps). We derive human-computer interaction requirements: clear scope-of-use messaging, prompt scaffolding, human-in-the-loop, privacy-preserving defaults, and explicit escalation/hand-off pathways.