Skip to content
Modelwire
Subscribe

Study finds 41 LLMs give different mental health advice based on who's asking

Source published ·Modelwire updated

Original coverage: arXiv cs.CL ↗·How Modelwire adds context

The development

Researchers audited 41 language models to assess how they recommend mental health interventions across three contexts: clinician-to-patient, caregiver-to-patient, and direct user queries. The study analyzed over 35,000 responses across clinical vignettes and found significant behavioral variance by context, with models offering consultation-style guidance to clinicians while providing hedged, simplified advice to lay users. This work exposes a critical gap in LLM consistency and safety for high-stakes domains, revealing that model recommendations diverge based on framing rather than clinical evidence alone. The findings underscore the need for rigorous auditing before deploying language models in healthcare workflows.

Modelwire’s AI-generated summary of coverage from arXiv cs.CL.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

The study's core finding isn't just that models give different advice to different audiences (that's expected), but that they do so inconsistently within the same clinical scenario. The variance appears driven by conversational framing rather than clinical evidence, suggesting models lack stable decision-making anchors in mental health domains.

This work extends the safety evaluation framework established in the SAFE-Teen study from September, which tested chatbot responses across adolescent mental health conversations using clinical criteria. Where that work measured failure modes in teen-specific contexts, this audit broadens the lens to show how the same models behave differently when addressing clinicians versus patients. The finding also connects to the clinical documentation loop problem identified last week: if models can't maintain consistent reasoning within a single domain, downstream systems that consume their outputs inherit that instability. Together, these papers suggest the issue isn't isolated chatbot failures but systematic brittleness in how LLMs handle high-stakes medical reasoning across different stakeholder roles.

If the researchers release a follow-up showing that fine-tuning models on clinically-grounded decision trees reduces context-dependent variance below a specified threshold (say, under 15% recommendation divergence on identical cases), that would indicate the problem is remediable before deployment. If no such follow-up appears within six months, it signals the variance may be structural to how these models weight conversational signals versus clinical evidence.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·arXiv cs.CL

    Large-scale safety audit finds chatbot failures shift with teen responses

    Researchers conducted a large-scale safety evaluation of 18 chatbot APIs in simulated mental health conversations with adolescents, testing 32,555 scenarios across 106 seed situations. The study, informed by clinical expertise, measured how chatbot responses vary based on subtle shifts in teen inputs and system prompting strategies, using 22 safety criteria tied to documented harms. This…

    Read Modelwire coverage →Original source ↗

MentionsLanguage models (41 models audited) · Mental health interventions · Clinical vignettes

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Curbside Chatbots: Auditing the Mental Health Interventions Large Language Models Recommend to Clinicians, Caregivers, and Patients”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

LLMs overestimate cancer survival, miss frailty in oncology reasoning

arXiv cs.CL·

GPT-6 tested on clinical model auditing without code execution

Six LLM families audited for racial bias in pain management recommendations

arXiv cs.CL·