JMIR Publications

ABSTRACT

Background:

Large language models (LLMs) are increasingly being integrated into clinical decision support, ranging from documentation aids to triage and patient communication. However, their behavior in ethically ambiguous medical situations is poorly characterized. Variability in how different LLMs balance approval against refusal of a proposed action could materially shape the recommendations clinicians receive, yet no widely shared benchmark systematically quantifies this variability across contemporary model families.

Objective:

This study aimed to develop an expert-curated benchmark of medical ethical dilemmas and to quantify model-to-model variability in binary ethical recommendations across five state-of-the-art LLMs, with the goal of supporting routine auditing and transparent reporting of LLM ethical decision tendencies.

Methods:

We constructed 60 original medical ethical dilemmas spanning four domains (Clinical Practice, Technology and AI, Research Ethics, and Public Health and Policy). Each case required a binary (Yes/No) recommendation with a brief justification. Three domain experts annotated every case by difficulty (Easy/Medium/Hard) and ethical principle (Autonomy, Beneficence, Non-maleficence, Justice). Five LLMs—Gemini 2.5, DeepSeek-R1-Distill-Qwen-1.5B (DeepSeek-R1-1.5B), Llama-3-8B, Claude-4-Sonnet, and GPT-5—were queried under a fixed zero-temperature prompt. Heterogeneity in positive response rates across models was assessed with Cochran’s Q test; pairwise comparisons used generalized estimating equations (GEE) with a binomial-logit link to account for within-case clustering. Inter-model agreement was summarized with pairwise Cohen’s κ and overall Fleiss’ κ. A 4:1 discordance analysis identified instances in which a single model deviated from the unanimous judgment of the remaining four.

Results:

Across 300 case–model decisions with no missing data, positive response rates differed markedly across models (Cochran’s Q=66.91, P<.001), ranging from 0.85 (95% CI: 0.75–0.93) for DeepSeek-R1-1.5B to 0.27 (95% CI: 0.17–0.38) for Claude-4-Sonnet. Inter-model agreement was only fair overall (Fleiss’ κ=0.31, 95% CI 0.23–0.40) and lowest for Hard cases (κ=0.24) and Public Health and Policy scenarios (κ=0.22). GEE comparisons confirmed substantially higher odds of approval for DeepSeek-R1-1.5B than Claude-4-Sonnet (odds ratio [OR] 15.58, 95% CI 6.71–36.19). DeepSeek-R1-1.5B was the sole outlier in 19 of 29 discordant 4:1 splits, with 18 of these 19 being “Yes-when-others-said-No”. Claude-4-Sonnet contributed 6 discordant “No” responses, whereas GPT-5 was never the sole outlier.

Conclusions:

State-of-the-art LLMs disagree substantially on the same medical ethical dilemmas, and the choice of model can materially influence the recommendation a clinician receives in ethically sensitive contexts. We argue that documentation of alignment tendencies—such as approval-versus-refusal bias across domains—should accompany medical AI release alongside accuracy metrics, and that LLM outputs in ethically sensitive clinical scenarios should be treated as decision support that requires human oversight. Our openly released benchmark provides a reusable foundation for routine auditing of LLM ethical recommendation behavior. Clinical Trial: not applicable