The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models
Tassallah Abdullahi, Shrestha Ghosh, Hamish S Fraser, Daniel León Tramontini, Adeel Abbasi, Ghada Bourjeily, Carsten Eickhoff, Ritambhara Singh
2026-01-12
Summary
This research investigates how giving large language models (LLMs) a 'persona' – like acting as a doctor or nurse – affects their performance and safety when dealing with medical decisions.
What's the problem?
We often assume that giving an LLM a professional persona automatically makes it better and safer at medical tasks. However, it wasn't clear if this was actually true, especially in real-world, high-pressure clinical situations. There was a need to understand *how* these personas influence LLM behavior, and if improvements in some areas might come at the cost of performance in others.
What's the solution?
The researchers tested different personas (like emergency room doctor versus nurse) and interaction styles (being bold versus cautious) with several LLMs on tasks like deciding how urgently a patient needs care and identifying potential safety risks. They didn't just look at whether the LLM got the right answer, but also how confident it was in its answer and whether its reasoning seemed safe. They compared the LLM’s responses to both other LLMs and actual human doctors to see how well the personas worked.
Why it matters?
The study found that personas don't guarantee better or safer performance. In fact, a medical persona could *improve* performance in critical care but *hurt* it in routine care. This shows that personas are more like starting points or biases for the LLM, creating trade-offs rather than automatic expertise. It highlights the need for careful evaluation and understanding of how these personas affect LLM behavior before relying on them for important medical decisions, and that even experts disagree on the quality of the LLM's reasoning.
Abstract
Persona conditioning can be viewed as a behavioral prior for large language models (LLMs) and is often assumed to confer expertise and improve safety in a monotonic manner. However, its effects on high-stakes clinical decision-making remain poorly characterized. We systematically evaluate persona-based control in clinical LLMs, examining how professional roles (e.g., Emergency Department physician, nurse) and interaction styles (bold vs.\ cautious) influence behavior across models and medical tasks. We assess performance on clinical triage and patient-safety tasks using multidimensional evaluations that capture task accuracy, calibration, and safety-relevant risk behavior. We find systematic, context-dependent, and non-monotonic effects: Medical personas improve performance in critical care tasks, yielding gains of up to sim+20% in accuracy and calibration, but degrade performance in primary-care settings by comparable margins. Interaction style modulates risk propensity and sensitivity, but it's highly model-dependent. While aggregated LLM-judge rankings favor medical over non-medical personas in safety-critical cases, we found that human clinicians show moderate agreement on safety compliance (average Cohen's κ= 0.43) but indicate a low confidence in 95.9\% of their responses on reasoning quality. Our work shows that personas function as behavioral priors that introduce context-dependent trade-offs rather than guarantees of safety or expertise. The code is available at https://github.com/rsinghlab/Persona\_Paradox.