Auditing Value Pluralism in Medical AI Language Models
A recently published paper on arXiv (2605.18738) introduces a novel framework for evaluating value pluralism within medical AI. This study explores the fundamental pluralism found in medicine, where principles such as autonomy, beneficence, nonmaleficence, and justice frequently clash, with effective clinical practice balancing these conflicts according to each patient's values. The authors offer a benchmark of ethical dilemmas verified by clinicians and a method for attributing value priorities based on AI decisions. Their examination of advanced models indicates variability in values among physicians and the presence of Overton pluralism in reasoning; however, the decisions made by individual models tend to be nearly deterministic across different samples and semantic changes, lacking the essential pluralism.
Key facts
- arXiv paper 2605.18738 introduces a framework for auditing value pluralism in medical AI.
- Medicine is inherently pluralistic with conflicting principles like autonomy, beneficence, nonmaleficence, and justice.
- The framework includes a benchmark of clinician-verified ethical dilemmas.
- An attribution method recovers value priorities directly from AI decisions.
- Frontier models show physician-level value heterogeneity.
- Models discuss competing values in reasoning (Overton pluralism).
- Individual model decisions are near-deterministic across repeated sampling and semantic variations.
- The research highlights a failure of current AI to reflect necessary pluralism in clinical ethics.
Entities
Institutions
- arXiv