Sensing but not alerting: ChatGPT mental health triage gaps in simulated psychodermatology conversations
Research letter testing whether ChatGPT recognises mental health concerns and recommends appropriate referral during simulated multi-turn psychodermatology conversations. Fifty first-person narratives drawn from published psychodermatological case reports were fed section by section to GPT-5 with conversational context preserved, three replicates each, and three board-certified psychiatrists blinded to the study aims rated each narrative's referral need. Analysis focused on the 28 narratives all three psychiatrists rated as needing at least therapy-level referral, and reports a large gap between the model noticing a mental health concern and acting on it.
Publisher
JAAD International (Elsevier, for the American Academy of Dermatology)
Published
25 Jun 2026
Added
3 days ago
Key Findings
- Across the 28 clinician-validated cases, 95.2% ± 5.5% of conversations produced a dermatology follow-up recommendation but only 52.4% ± 5.5% produced a mental health follow-up recommendation; 71.4% ± 3.6% flagged a mental health concern without necessarily referring
- In the 9 cases unanimously rated as definitely needing psychiatric referral, 77.8% of simulations expressed mental health concern but only 48.1% recommended mental health follow-up
- Delusional cases fared worse on referral than non-delusional ones (45.8% ± 3.6% versus 61.1% ± 12.7%)
- In 7 delusional-infestation cases, 57.1% of conversations advised the user to collect specimens for a future visit, advice the authors note may validate the infestation premise and reinforce compulsive evidence-gathering
- The authors report observing similar triage gaps in Gemini 3
Methodology Notes
Simulation study, not a study of real patients. Narratives were authored by the research team from published case reports, and the authors note that the drafting process may attenuate mental health signals, which is why analysis was restricted to the 28 clinician-validated cases. Model tested was gpt-5-2025-08-07 via API, three replicates per narrative, with three blinded board-certified psychiatrists providing the referral ground truth. Short-format research letter (three pages, five references) with no abstract; NSF-funded (2125872); the authors disclose using ChatGPT-5 Thinking for writing clarity. Date precision: Crossref records the article as pages 166-168 of the August 2026 issue with a deposit date of 2026-06-25 and no separate published-online date-part; PubMed records 2026 Jun 25. The 2026-06-25 date is used here and the issue is August 2026. Verification route: ScienceDirect and the PMC PDF are both bot-blocked from this box; the publisher full text was read through the r.jina.ai text proxy, with the PMC landing page (PMC13400423) and the Crossref DOI record confirming journal, pages and CC BY licence.
Sources
Authors
Mohammad Iqbal Nouyed, Lauren E. Kozlowski, Evelyn Shue, Alexsandra E. Smith, Dilip N. Chandran, Daniel E. Elswick, Michael S. Kolodney, Wanhong Zheng, Gangqing Hu
Tags
Cite This
APA
Mohammad Iqbal Nouyed et al. (2026). Sensing but not alerting: ChatGPT mental health triage gaps in simulated psychodermatology conversations. JAAD International (Elsevier, for the American Academy of Dermatology). https://www.jaadinternational.org/article/S2666-3287(26)00117-3/fulltext
Related Insights
Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment
Psychiatric Services (American Psychiatric Association); RAND-led author team · 26 Aug 2025
A clinically validated framework for auditing AI chatbot behavior in mental health interactions
Nature Medicine · 7 Aug 2026
Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots
BJPsych Open (Cambridge University Press / Royal College of Psychiatrists) · 11 Jun 2026