Research map / Large language models / AI and society
What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Audits value pluralism of 12 frontier LLMs on 50 clinician-verified principlist dilemmas, using a logit-based value attribution method and comparison to a 20-physician panel.
Problem
Medical LLMs make value-laden clinical recommendations (autonomy, beneficence, nonmaleficence, justice), but their ethical priorities and whether they reflect the pluralism of physician opinion have not been systematically measured.
Approach
A clinician-verified benchmark of binary clinical dilemmas with principlist value tags, plus a value-attribution method that infers value priority distributions from observed decisions (logit regression on value-difference vectors with temperature-scaled softmax), used to compare LLM decisions against physician consensus and inter-physician variation.
Key results
- 11 of 12 models have median decision entropy of zero across 10 repeated samples; per-model correlation with physician disagreement is near zero (mean rho -0.021, all p>0.17).
- Most models fall within inter-physician JSD range, but GPT 5.2, Grok 4, and Sonar Pro exceed the 95th percentile, underweighting autonomy (6.1% to 12.8% vs physician consensus 44.4%).
- Choice-balanced Overton coverage is high (mean OvCov 0.86), while emphasis is lower (OvEmph 0.61); LLM ecosystem diversity (0.0916) does not differ significantly from physician diversity (0.1089).
Surprising findings
- Models are near-deterministic across repeated sampling even on cases where physicians split evenly, with no correlation to physician disagreement.
- Frontier model ecosystem diversity in value profiles is statistically indistinguishable from that of a physician panel, so the monoculture risk lies at deployment rather than across models.
Limitations
- Physician sample of 20 is not representative; results are calibration rather than population claims.
- Forced-choice binary design limits ecological validity; English-language, Western bioethics focus.
- Logit attribution assumes linear separability of value contributions; snapshot of model versions.
Experimental setup
Datasets: biomedical ethics literature seeds
Benchmarks: 50-case clinical ethics dilemma benchmark
Metrics: decision entropy Spearman correlation with physician entropy Jensen-Shannon divergence OvCov OvEmph likelihood ratio test flip rate
Baselines: physician panel consensus
Builds on: pluralistic alignment (Sorensen et al.) principlism OvertonScore
Closest papers by meaning
- Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties (AAAI 2024 · similarity 0.85)
- Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning (ICLR 2026 · similarity 0.85)
- Position: A Roadmap to Pluralistic Alignment (ICML 2024 · similarity 0.85)
- Why Specialist Models Still Matter: A Heterogeneous Multi-Agent Paradigm for Medical Artificial Intelligence (ICML 2026 · similarity 0.84)
- Position: Beyond Assistance – Reimagining LLMs as Ethical and Adaptive Co-Creators in Mental Health Care (ICML 2025 · similarity 0.82)
- Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.