November 18, 2026
Health news, explained with evidence
LIVE

Study shows decision-style prompts shift clinical thresholds in large language models

🧠 AI Analysis 2 min read 👁 4

A recent evaluation by researchers published in Nature Health Sciences shows that system prompts mimicking different physician personas can significantly alter clinical action thresholds in large language models under fixed evidence.

Study shows decision-style prompts shift clinical thresholds in large language models🧠AI Analysis

Photo: Nature Health Sciences

Key points

  • System prompts shifting clinical action thresholds
  • Twenty open-weight models tested on 5,000 total cases
  • Affirmative rates shifted by 9.5 percentage-point swing

As the deployment of clinical large language models shifts from theoretical feasibility to practical liability, researchers investigated whether decision-style system prompts alter clinical action thresholds while keeping clinical facts constant. The evaluation tested twenty open-weight models across 2,500 simulated emergency department vignettes and 2,500 MIMIC-IV-Note discharge summaries. By crossing three ethical orientations—duty-based, care-based, and utilitarian—with three cognitive styles—intuitive, integrative, and analytic—the team defined nine distinct physician personas to assess binary decision items covering safety, autonomy, treatment, resource use, and follow-up.

Under baseline prompting, the models answered affirmatively to 42.8% of decisions. However, introducing persona prompts shifted these affirmative rates from 36.9% to 46.4%, creating a 9.5 percentage-point swing even though the underlying clinical evidence remained fixed. The researchers observed that these effects were particularly pronounced in categories involving patient autonomy and treatment recommendations, remaining consistent across the tested corpora. Furthermore, the susceptibility to prompt framing varied among models from 4.9 to 16.1 points, and neither medical fine-tuning nor larger model size provided consistent protection against these shifts.

What to know: The findings suggest that system prompts act as a policy-setting layer rather than a simple communication layer, meaning prompt configurations in healthcare settings require careful governance. Because these models lack human grounding, lived experience, or longitudinal feedback, changes in framing can systematically alter output thresholds. Readers should consult qualified healthcare professionals for all personal medical decisions and policy interpretations.

Source Nature Health Sciences — nature.com/articles/s44401-026-00121-z ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.