Do AI doctors have hidden ethical biases?

This paper audits whether medical language models reflect the ethical pluralism of human physicians—the ability to weigh competing principles like autonomy, beneficence, and justice based on patient values. The authors created a benchmark of clinician-verified ethical dilemmas and developed an attribution method to extract value priorities from model decisions. While models can discuss competing values in their reasoning, their actual decisions are near-deterministic and fail to match physician diversity. Most models' priorities fall within normal physician variation, but some systematically underweight patient autonomy. The finding suggests that deploying a single LLM without attention to its embedded values could amplify narrow ethical perspectives at scale across populations.