What Does the AI Doctor Value? Auditing Pluralism i... | AI Research

What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models
Medicine is a field where there is rarely one "correct" answer to complex ethical dilemmas. Physicians often disagree on how to balance competing principles like patient autonomy, beneficence, nonmaleficence, and justice. While human doctors are expected to navigate these tensions in collaboration with their patients, the ethical priorities of artificial intelligence (AI) models remain largely unexamined. This paper introduces a new framework to audit how large language models (LLMs) handle these ethical trade-offs, assessing whether they reflect the diverse perspectives found in clinical practice or if they risk imposing a singular, rigid ethical stance.

Auditing Ethical Decision-Making

To evaluate how AI models handle moral conflict, the researchers created a benchmark consisting of 50 clinical dilemmas. Each case was carefully crafted to force a choice between two mutually exclusive recommendations, where selecting one option inherently prioritizes certain ethical values over others. By using a blinded review process involving expert physicians, the team ensured the cases were clinically realistic and captured genuine ethical tension. They then tested various frontier LLMs by asking them to provide medical advice for these scenarios, using a specialized attribution method to uncover the "value profile" or ethical priorities hidden behind each model's decisions.

Consistency Versus Pluralism

The study found that while individual physicians often disagree with one another, AI models are remarkably consistent. When faced with the same clinical dilemma, a model will almost always make the same decision, showing near-zero decision entropy. This consistency is not necessarily a sign of accuracy; rather, it indicates that models act as deterministic decision-makers. While human doctors represent a wide spectrum of ethical viewpoints, the models fail to mirror this "distributional pluralism." Even when models discuss competing values in their reasoning—a trait the authors call "Overton pluralism"—their final decisions remain rigid, suggesting that their internal logic is driven by fixed, systematic value preferences.

The Risk of a Deployment Monoculture

The researchers discovered that most frontier models hold value priorities that fall within the natural range of human physician variation. However, a few models significantly deviate from the physician consensus, particularly by underweighting patient autonomy. This finding is critical because, in a real-world setting, a patient typically interacts with only one AI model. If that model consistently prioritizes certain values over others, it could systematically skew the advice given to every patient it serves. Without intentional efforts to balance these ethical perspectives, the widespread deployment of a single LLM risks replacing the healthy, pluralistic nature of clinical practice with a "deployment monoculture" that lacks the flexibility and diversity of human medical judgment.