Curbside Chatbots: Auditing the Mental Health Interventions Large Language Models Recommend to Clinicians, Caregivers, and Patients
Organizations: Division of Digital Psychiatry, Beth Israel Deaconess Medical Center, Boston, Massachusetts, United States · Department of Psychiatry, Beth Israel Deaconess Medical Center, Boston, Massachusetts, United States · Psychiatry and Mental Health Department, School of Medicine, Pontificia Universidad Javeriana, Bogotá, Colombia · Department of Psychiatry and Psychotherapy, Charité – Universitätsmedizin Berlin, Campus Benjamin Franklin, Berlin, Germany
Abstract
Users are asking language models mental health questions and are receiving answers. We audited what 41 language models recommend to users when the same clinical information reaches them from three footings: a clinician describing a patient, a caregiver describing a loved one, or a person describing themselves. Across 19 case vignettes and five questions, we collected 35,055 free-text responses, extracted every intervention named with a validated judge, and had 29 models rate the full option lists on a seven-point Likert scale for the same cases. Models largely recommend what they rate well, with recommendation behavior differing by footing, and many models offering clinicians a consultation-style response and lay users hedged, simplified advice. When asked about recommendations for a specific intervention type, models more often named options they themselves rated below zero as initial choices, usually with conditions attached. We will release the instrument, code, and full corpus on publication as a living audit.
Figures & tables
| Footing | Writing Style | Key Characteristics | Example |
|---|---|---|---|
| Clinician | Telegraphic clinical chart style | Concise, clinical terminology, abbreviated descriptions, diagnosis named | Moderate major depressive disorder: low mood, anhedonia, low energy, and poor concentration most days x 4 months. Reduced functioning at work. What would you recommend for this patient? |
| Caregiver | Plain English prose | Natural, conversational language describing a loved one’s condition | My wife has been really down for the past four months or so. She has no energy, she’s stopped enjoying the things she used to love, she can’t concentrate, and she’s barely getting through work. What would you recommend for them? |
| Patient | Casual, informal writing | Quick-typing style with missing apostrophes, inconsistent capitalization, misspellings, and run-on sentences | Ive been really depressed for like 4 months now. No energy, cant enjoy anything, cant focus on anything, and im barely getting thru work. What would you recommend for me? |
| Arm | Models | Cells per model | Total cells |
|---|---|---|---|
| Open-ended generation | 41 | 855 | 35,055 |
| Structured rating | 29 | 684 | 19,836 |
Medical LLMs
Medical large language models are AI systems adapted to clinical and biomedical data for health reasoning, information extraction, and communication.
See all 428 papers