Variation in Medical Ethics Recommendations Across Large Language Models: Case-Based Audit Study

@article{mitsuyama2026variation,
  title = {{Variation in Medical Ethics Recommendations Across Large Language Models: Case-Based Audit Study}},
  author = {Yasuhito Mitsuyama and Ko Watanabe and Shannon Walston and Hirotaka Takita and Kenichi Saito and Shoya Ishimaru and Daiju Ueda},
  journal = {JMIR AI},
  year = {2026},
  month = oct,
  doi = {10.2196/102713},
  url = {https://doi.org/10.2196/102713},
  note = {CC BY 4.0},
}

TY  - JOUR
TI  - Variation in Medical Ethics Recommendations Across Large Language Models: Case-Based Audit Study
AU  - Yasuhito Mitsuyama
AU  - Ko Watanabe
AU  - Shannon Walston
AU  - Hirotaka Takita
AU  - Kenichi Saito
AU  - Shoya Ishimaru
AU  - Daiju Ueda
PY  - 2026
DA  - 2026/10/19
JO  - JMIR AI
DO  - 10.2196/102713
UR  - https://doi.org/10.2196/102713
AB  - 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 of 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 a standardized case-based audit set of medical ethics dilemmas and quantify variation in recommendation direction, model-reported confidence, within-model stability, and between-model agreement across 5 selected contemporary LLMs.  Methods: We constructed 60 original medical ethics dilemmas spanning 4 domains. Three researchers independently annotated case difficulty and ethical principles before reaching consensus. Five primary models—GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro, full-size DeepSeek-R1, and Meta Llama-3-8B-Instruct—were each evaluated 3 times per case using a fixed prompt and temperature 0. DeepSeek-R1-Distill-Qwen-1.5B was evaluated separately as an exploratory small-model comparator. The primary outcome was the binary yes or no answer; model-reported confidence and reason were also collected. Model differences were assessed using a case-clustered generalized estimating equation (GEE), and between-model agreement and within-model stability were summarized.  Results: Of 900 planned primary panel outputs, 897 (99.7%) had valid binary answers. Trial-level yes rates were 44.4% (80/180) for DeepSeek-R1, 42.4% (75/177) for Claude Sonnet 4.5, 38.9% (70/180) for GPT-5, 38.9% (70/180) for Gemini 2.5 Pro, and 16.7% (30/180) for Llama-3-8B. The overall model effect was significant (Wald χ24=23.34; P<.001); after Holm adjustment, only comparisons involving Llama-3-8B were significant. Across 59 complete-panel cases, raw agreement was 79.0% (466/590 model-pair comparisons; 95% bootstrap CI 72.9%-84.7%), and Fleiss κ was 0.55 (95% bootstrap CI 0.43-0.66). Within-model response stability across 3 runs ranged from 85.0% (51/60) to 100.0% (59/59 and 60/60).  Conclusions: Selected contemporary LLMs showed model-dependent variation but moderate overall agreement in binary recommendations for medical ethics cases. These results characterize recommendation behavior under a fixed, time-specific protocol and do not establish ethical correctness or the effects on clinicians or patients. Reporting model version, recommendation tendency, confidence, and repeatability may improve transparency when such systems are evaluated for ethically sensitive applications.
SN  - 2817-1705
ER  -