The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models

Quick Overview

Persona conditioning, the common practice of using a single system prompt like "You are an ED physician" to ensure safety in clinical LLMs, creates a "persona paradox" where the persona significantly improves performance in high-acuity emergencies but actively degrades performance and introduces dangerous biases in lower-acuity, routine cases, necessitating task-conditional evaluation.

Key Points: Medical personas consistently and significantly improve performance in high-acuity emergency cases, boosting accuracy by up to 20 percentage points and dropping Expected Calibration Error (ECE) by about 20 points. The same medical personas frequently degrade performance on routine primary care scenarios, causing accuracy drops of about 10 percentage points because the persona's built-in bias toward escalation overtriage low-risk cases. The ED physician persona possesses a powerful bias toward escalation, forcing the model to overtriage by acting as a fixed behavioral prior optimized only for finding life-threatening emergencies. Stylistic prompts, such as testing bold versus cautious styles for the ED physician, induce shifts in risk propensity that are completely inconsistent across different tested LLMs, rendering them unreliable as safety controls. Automated LLM judges consistently preferred medical personas over non-medical ones for safety and helpfulness, but human clinicians reported extremely low confidence (95.9% of ratings) when judging the justification quality or reasoning behind triage decisions. The study mandates task-conditional evaluation, concluding that one cannot deploy a single doctor persona across an entire hospital system because the persona must precisely match the context of the task. A low Consistency Rate (CR) indicates the persona is making the model 'lie' by painting a clinical facade over an internally different or uncertain decision, rather than truly changing the underlying reasoning.

Context: The discussion centers on a research paper investigating the effectiveness and safety implications of using 'persona conditioning'—a simple, one-sentence system prompt (e.g., 'You are an ED physician')—to steer Large Language Models (LLMs) in high-stakes clinical settings like decision support. The researchers designed an experimental framework to test this common assumption by comparing specific medical roles (ED physician, ED nurse) against non-medical baselines across two domains: high-acuity clinical triage and general patient safety compliance, using metrics beyond simple accuracy, specifically Expected Calibration Error (ECE) and risk posture analysis.