Personas Shift Clinical Action Thresholds in Large Language Models
Eyal Klang, Alon Gorenshtein, Mahmud Omar, Girish N. Nadkarni · medRxiv · 2026
Background and aims: Clinical LLM deployment is shifting from feasibility to liability, while current guidance largely treats model behavior as a control problem. We tested whether decision-style system prompts shift clinical action thresholds when clinical facts are held constant, and whether these shifts are consistent across settings and models. Methods: We defined nine physician personas by crossing three ethical orientations (duty-, care-, utilitarian) with three cognitive styles (intuitive, integrative, analytic). Twenty open-weight LLMs were evaluated on 2,500 simulated ED vignettes and 2,500 MIMIC-IV-Note discharge summaries. For each text, models answered five binary decision items (safety, autonomy, treatment, resource use, follow-up). Each condition was repeated ten times, yielding 5,000,000 total decisions. Results: Under baseline prompting, models answered "Yes" to 42.8% of decisions. Persona prompts shifted affirmative rates from 36.9% to 46.4%, a 9.5-percentage-point swing under fixed clinical evidence. Effects were largest in autonomy and treatment and were consistent across corpora (85.7% directional agreement; r = 0.82 for effect sizes). Susceptibility varied by model (4.9-16.1 points), with no consistent protection from medical fine-tuning or model size. Conclusions: Decision-style system prompts reliably change clinical action thresholds in LLMs under fixed evidence. Prompting is a policy-setting layer, not just a communication layer, and should be treated as a first-class deployment configuration.