Skip to main content
Peer-reviewed Authoritative

Between Help and Harm: An Evaluation Study of Mental Health Crisis Handling by Large Language Models

Peer-reviewed study introducing a taxonomy of six clinically-informed mental-health crisis categories, an evaluation dataset of over 2,000 user inputs drawn from twelve public conversational datasets, and an expert protocol for rating response appropriateness and safety. Assesses how leading LLMs handle crisis conversations.

Publisher

JMIR Mental Health

Published

11 Jun 2026

Added

today

Key Findings

  • Releases a six-category crisis taxonomy and an evaluation dataset of 2,000+ inputs aggregated from twelve datasets
  • A non-negligible share of LLM responses were inappropriate or harmful, worst for self-harm and suicidal ideation
  • Safety failures tracked alignment quality more than raw model scale, and models were weakest on indirect distress signals

Methodology Notes

Peer-reviewed, JMIR Mental Health 2026, vol. 13, article e88435 (11 June 2026), DOI 10.2196/88435. Published successor to arXiv preprint 2509.24857. Benchmark built by aggregating twelve existing datasets; expert-rated response protocol. mental.jmir.org is JS-rendered to fetchers; DOI/date confirmed via Crossref.

Authors

Adrian Arnaiz-Rodriguez, Miguel Baidal, Erik Derner, Jenn Layton Annable, Mark Ball, Mark Ince, Elvira Perez Vallejos, Nuria Oliver

Tags

jmircrisis-handlingbenchmarktaxonomypublished-version

Cite This

APA

Adrian Arnaiz-Rodriguez et al. (2026). Between Help and Harm: An Evaluation Study of Mental Health Crisis Handling by Large Language Models. JMIR Mental Health. https://mental.jmir.org/2026/1/e88435