55/100

Moderate evidence

ProblemPublished recordHealthProblem Space

Personas shift clinical action thresholds in large language models

Truvace record
TRV-2026-2937raw data (JSON)
Version
v1Published
First published
Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
sha256
3aa1c9388440…check it

Clinical large language model (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. We defined nine physician personas by crossing three ethical orientations (duty-based, care-based, utilitarian) with three cognitive styles (intuitive, integrative,…

The problem

In a test of 20 open-weight language models on 5,000,000 decisions, doctor-persona prompts moved the share of 'Yes' answers from 36.9% to 46.4% (a 9.5-point swing) with the clinical facts held constant.

What this doesn’t fix

The decisions were made on simulated emergency vignettes and retrospective discharge summaries by open-weight models, so the study does not measure effects on actual patient care.

Sources

1 on file
  1. npj Health Systemsdoi:10.1038/s44401-026-00121-z
    Peer-reviewed2026-11-14

Cite this record

v1

Truvace Impact Record TRV-2026-2937, v1: “Personas shift clinical action thresholds in large language models.” Truvace, 2026-11-15. /record/TRV-2026-2937 (accessed on the day you copy it). sha256 3aa1c9388440321f…

Change history

1 version

Every change made to this record since it was first published, and why. Nothing has changed since it went up.

  1. Publishedv13aa1c9388440…

    Certified into the record

Check this record hasn’t been altered

Check this record hasn’t been altered
How to check it yourself, without trusting this page

Download any version from /api/record/TRV-2026-2937 and hash its canonical field yourself, for example with shasum -a 256. You should get its content_hash. Each version’s text also ends with prev:and the hash of the version before it (version 1 points to 64 zeros). If anyone had changed a single character of any version after it was saved, those hashes wouldn’t line up.