Decision-style prompts changed how clinical LLMs answered the same case
Clinical large language models are moving out of the laboratory and into workflows where a yes-or-no recommendation can shape triage, follow-up, treatment, and the handling of patient autonomy. This study tested a simple but consequential question: when the clinical facts stay fixed, do persona-based instructions change the threshold at which a model decides to act?
They do.
Across 20 open-weight models, nine physician personas produced measurable swings in clinical recommendations, even though the underlying vignettes and discharge summaries did not change. Under baseline prompting, models answered “Yes” to 42.8% of decisions. When the system prompt imposed a physician persona, affirmative responses shifted from 36.9% to 46.4%, a 9.5 percentage-point spread under the same evidence. In practical terms, the model’s policy layer was not neutral; it nudged the same clinical scenario toward more permissive or more restrictive action.
Five million decisions, two clinical corpora, one consistent signal
The design was intentionally broad. Investigators built nine personas by crossing three ethical orientations — duty-based, care-based, and utilitarian — with three cognitive styles — intuitive, integrative, and analytic. Those prompts were then tested across 2,500 simulated emergency department vignettes and 2,500 MIMIC-IV-Note discharge summaries. For each text, models answered five binary items covering safety, autonomy, treatment, resource use, and follow-up. Each condition was repeated ten times, producing 5,000,000 total decisions.
The scale matters because it makes the signal hard to dismiss as prompt noise. The same clinical record could move a model across a meaningful decision boundary simply by changing the stance embedded in the system instructions. In a healthcare setting, a small shift in the rate of “yes” responses can translate into more tests ordered, more treatment endorsed, more follow-up advised, or more patient preference overridden depending on how the item is framed. When the model is functioning as an agentic support tool, threshold drift is not abstract. It changes downstream action.
Autonomy and treatment moved the most
The strongest persona effects appeared in autonomy and treatment decisions. Those domains are where clinical language models confront the tension between rule-following, beneficence, and deference to patient preference. A duty-based prompt may bias the system toward obligation and consistency, while a care-based prompt may weight relationship and context. A utilitarian prompt can pull toward outcomes and aggregate benefit, which may justify action in one setting and restraint in another. The result is not a philosophical flourish; it is a measurable shift in whether the model recommends intervention.
One sentence captures the operational risk: the same case can receive a different action threshold without any change in its facts.
The investigators also found strong consistency across the two corpora. Directional agreement reached 85.7%, and effect sizes correlated at r = 0.82 between emergency department vignettes and discharge summaries. That level of concordance argues against the shifts being a peculiarity of one dataset or one clinical context. A model that is steerable in one domain appears steerable in another, even when the source material differs in structure and clinical stage.
Model-to-model variation was wide, and fine-tuning did not reliably protect against it
Susceptibility varied substantially across models, from 4.9 to 16.1 percentage points. Some systems were relatively stable, while others were highly responsive to persona framing. No consistent protection emerged from medical fine-tuning or model size. Larger models were not automatically more resistant, and clinical tuning did not reliably immunize a system against prompt-based shifts in action threshold.
That finding carries direct deployment implications. Health systems often assume that a model trained on medical data is inherently safer or more predictable in clinical use. These data suggest a more complicated reality: once a system prompt defines an ethical or cognitive stance, the response surface can shift in ways that are not neutralized by training pedigree alone. In other words, model selection is not enough. Prompt governance becomes part of the clinical safety case.
Why this happens matters for bedside use
Large language models do not possess moral commitments, but they do reproduce regularities in the text patterns they were trained on and in the instruction structure placed in front of them. A persona prompt functions like a policy lens. It tells the model which kinds of answers should sound more acceptable, more cautious, more obligation-driven, or more outcome-driven. When the same case is presented through a different lens, the model’s action threshold moves accordingly.
For clinicians, that means the prompt is not just a communication aid. It is a hidden configuration setting that can alter what the system recommends when the evidence is held constant. In settings such as emergency care, discharge planning, and follow-up prioritization, that shift may affect triage urgency, the endorsement of intervention, or the handling of a patient’s stated preferences. The danger is not only incorrect output. It is inconsistent output that depends on invisible framing choices.
Governance has to treat prompting as a deployment variable
The study’s conclusion is direct: decision-style system prompts reliably change clinical action thresholds in LLMs under fixed evidence. That makes prompting a policy-setting layer rather than a simple interface layer. If a health system allows clinicians or vendors to alter persona instructions without oversight, it is effectively allowing changes to the model’s action threshold while keeping the clinical scenario unchanged.
This also reframes what safety review should examine. It is not enough to ask whether the model can summarize a chart or answer a question. Regulators and health systems will need to ask how the system prompt shifts the balance between action and restraint, and whether those shifts are acceptable in specific clinical tasks. The same holds for monitoring after deployment, where output audits should track not just accuracy but prompt sensitivity.
A final caution follows from the data itself: the prompt layer is small, but the effect is not. On a five-million-decision evaluation, a seemingly modest change in affirmative rate exposed a real instability in how clinical authority is expressed by LLMs.