A Risk-Formulation Framework and Candidate Screening Items for Dangerous Use of Large Language Models in Clinical Psychiatry
Abstract
Large language models (LLMs) and related conversational artificial intelligence (AI) systems have moved rapidly from novelty into ordinary use across education, work, social life, and mental-health-related help seeking. Members of the public now use these systems for advice, companionship, self-disclosure, emotional support, psychoeducation, and quasi-therapeutic interactions. Existing instruments assess problematic or dependent AI use, but they do not ask whether AI has entered a pathway toward serious harm by entrenching ideation, operationalizing capability, displacing human support, or amplifying harmful conduct. To our knowledge, no brief, clinically oriented psychiatric screener has yet been developed and validated specifically for this purpose. We define dangerous AI use as generative AI use that appears to increase the likelihood, feasibility, persistence, organization, or scale of serious harm to self or others. This paper describes LLMs as interactive capability amplifiers and quasi-social systems rather than informational technologies alone, situates their risk within the technical problems of alignment and behavioral stability, and proposes a four-dimensional clinical formulation: Ideational Entrenchment, Operationalized Capability, Relational Displacement, and Broader Amplification. We then introduce two candidate instruments awaiting formal validation: the STEPS-5, a brief screen for AI involvement in acute self-harm and suicide risk, and the HEAT-4, a research-oriented framework addressing broader harmful-use amplification. These tools are intended to open clinical inquiry rather than diagnose, predict behavior, or determine disposition. Ultimately, we argue for the need to avoid both over-pathologizing ordinary AI use and ignoring when AI has become meaningfully relevant to psychiatric safety.
Related
- An AI-Enabled Translational Drug Discovery Framework for NALCN Channelopathy: From High-Throughput Screening to Therapeutic Candidate
- The Meta-Awareness – Equanimity Framework of Self for Psychiatry: Phenomenological Insights from Psychosis and Mindfulness
- Formononetin as a Natural Product-Derived Candidate for Diabetic Microvascular Complications: Mechanistic Evidence, Pharmacokinetics, Formulation Strategies, and Translational Barriers
- Burnout- and Depression-Like Functional Collapse States — A Within-Person Biomarker Framework for Precision Psychiatry: State-Dependent Multisystem Regulation in High Processing Cost Phenotypes (HPCPh)
- Screening traditional Chinese medicine monomers for premature ovarian insufficiency: from network meta-analysis to the effects of a promising candidate on oxidative stress biomarkers
BibTeX
@misc{braz2026riskformulation,
title = {{A Risk-Formulation Framework and Candidate Screening Items for Dangerous Use of Large Language Models in Clinical Psychiatry}},
author = {Armando N. Braz and Punit Matta and Joshuah Muthee and Eren M. Veziroglu and Daniel Price},
howpublished = {PsyArXiv},
year = {2026},
month = oct,
doi = {10.31234/osf.io/a2b9z_v1},
url = {https://osf.io/preprints/psyarxiv/a2b9z_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A Risk-Formulation Framework and Candidate Screening Items for Dangerous Use of Large Language Models in Clinical Psychiatry
AU - Armando N. Braz
AU - Punit Matta
AU - Joshuah Muthee
AU - Eren M. Veziroglu
AU - Daniel Price
PY - 2026
DA - 2026/10/09
PB - PsyArXiv
DO - 10.31234/osf.io/a2b9z_v1
UR - https://osf.io/preprints/psyarxiv/a2b9z_v1/
AB - Large language models (LLMs) and related conversational artificial intelligence (AI) systems have moved rapidly from novelty into ordinary use across education, work, social life, and mental-health-related help seeking. Members of the public now use these systems for advice, companionship, self-disclosure, emotional support, psychoeducation, and quasi-therapeutic interactions. Existing instruments assess problematic or dependent AI use, but they do not ask whether AI has entered a pathway toward serious harm by entrenching ideation, operationalizing capability, displacing human support, or amplifying harmful conduct. To our knowledge, no brief, clinically oriented psychiatric screener has yet been developed and validated specifically for this purpose. We define dangerous AI use as generative AI use that appears to increase the likelihood, feasibility, persistence, organization, or scale of serious harm to self or others. This paper describes LLMs as interactive capability amplifiers and quasi-social systems rather than informational technologies alone, situates their risk within the technical problems of alignment and behavioral stability, and proposes a four-dimensional clinical formulation: Ideational Entrenchment, Operationalized Capability, Relational Displacement, and Broader Amplification. We then introduce two candidate instruments awaiting formal validation: the STEPS-5, a brief screen for AI involvement in acute self-harm and suicide risk, and the HEAT-4, a research-oriented framework addressing broader harmful-use amplification. These tools are intended to open clinical inquiry rather than diagnose, predict behavior, or determine disposition. Ultimately, we argue for the need to avoid both over-pathologizing ordinary AI use and ignoring when AI has become meaningfully relevant to psychiatric safety.
ER -