跳到正文
原文
Nature 健康护理主题·· 6 天前AI 评分52

不同决策风格设定使大语言模型临床行动阈值发生显著偏移

Personas shift clinical action thresholds in large language models

AI 导读

该研究评估20个大语言模型在相同临床信息下因系统提示所赋予的不同决策风格而产生的临床判断差异。9种 persona 覆盖三种伦理取向与三种认知风格的交叉,共产生约500万次二元决策。最保守的Care-Analytic persona使

正文

Introduction

LLMs in clinical settings are no longer a feasibility story; they’re a liability story. The question is shifting from “can it help?” to “what exactly are we deploying when we deploy it?”1,2 Health systems and regulators are issuing guidance on clinical AI3. These guidelines focus on safety checks, documentation, human oversight, and limits on model autonomy4,5,6. They treat LLMs as tools whose behavior can be managed through access control, monitoring, and review. Governance is framed as a control problem.

At the same time, evidence shows that LLMs do more than retrieve facts or apply rules. Throughout this paper, language about “stances,” “orientations,” and “decision styles” refers to observable regularities in model outputs rather than to intentions, beliefs, or psychological dispositions; these regularities are best read as statistical mimicry of patterns in the training data and in the prompt, not as evidence that models hold values or cognitive commitments. They adopt stable response patterns that depend on context and framing7. The same clinical case, phrased with different instructions or premises, can yield different recommendations in ways that are systematic and reproducible rather than random8,9. These patterns resemble “reasoning stances” more than noise. In practice, that stance shows up as an action threshold: the same evidence can push the model toward a more permissive “yes” posture or a more restrictive “no” posture.

Clinical medicine has long described such stances in human decision-making. Clinicians differ in ethical orientation (duty-based, outcome-focused, or care-based) and in cognitive style (intuitive, integrative, or analytic). These orientations shape thresholds for action, willingness to override patient preferences, and tolerance for uncertainty. In human clinicians, they are recognized and studied; in language models, they are present but largely unnamed and unmeasured.

Human clinicians develop their decision styles over years of training, supervision, and shared practice. Their ethical orientation and cognitive habits are shaped by teams, departments, and real interactions with patients. LLMs have none of this grounding. They lack lived experience, social context, and longitudinal feedback, and we have little empirical data on how they adopt or express decision styles. This gap makes it necessary to study how framing influences their clinical judgments.

We tested whether explicit decision-style instructions change what care LLMs recommend when the clinical facts are fixed10,11,12,13. We examined how large these shifts are, how consistent they are across clinical settings, and whether some models are more vulnerable than others14,15.

Methods

Study design

We evaluated how explicit decision-style instructions influence clinical judgments made by LLMs. We defined nine decision styles (“physician personas”) by crossing three ethical orientations (duty-based, care-based, utilitarian) with three cognitive styles (intuitive, integrative, analytic). The ethical dimension draws on established normative frameworks in medical ethics: deontological approaches emphasizing rules and professional obligations, care ethics prioritizing relational context and patient-centered values, and utilitarian reasoning focused on outcome optimization16,17,18,19. The cognitive dimension reflects dual-process theory, distinguishing rapid intuitive reasoning from deliberate analytic processing, with integrative approaches balancing both modes. This 3×3 factorial design captures major axes along which clinician decision-making varies while remaining tractable for evaluation. Each persona was implemented as a short system prompt specifying priorities, decision rules, and risk posture, without adding clinical content (full persona definitions and prompt templates appear in Supplementary Methods Section 7; persona-specific effects are detailed in Table S1). To make the design concrete, one representative system prompt, used for the Duty-Analytic persona, read in part: “You are a physician whose decisions are guided by professional duty and systematic analysis. Prioritize adherence to established guidelines and documented procedures. Weigh the available evidence step by step before recommending action and avoid intervening in the absence of a clear indication.” The eight remaining personas replaced only the italicized priorities and decision rules; clinical content, response format, and task instructions were identical across personas (verbatim prompts for all nine personas are listed in Supplementary Methods Section 7).

LLMs were tested on two corpora: (1) 2500 simulated emergency department (ED) vignettes generated using a fixed zero-shot pipeline, and (2) 2500 real discharge summaries from the MIMIC-IV-Note database. For each text, models answered binary questions in five domains: safety, autonomy, treatment, resource use, and follow-up, reflecting common clinical action points. The clinical information was constant across conditions; only the decision-style framing changed (Fig. 1).

Fig. 1: Flowchart of the study design.

Created in BioRender. Omar, M. (2026) https://BioRender.com/fzj1gqf.

Vignette and note preparation

Simulated ED Vignettes: We developed 2,500 ED vignettes representing 25 common reasons for visit based on CDC NHAMCS classifications. Vignettes were 80–120 words, included essential clinical details, and excluded diagnoses or disposition statements. Two emergency physicians independently reviewed a random 10% sample for clinical realism; disagreements were resolved by consensus. Details of the generation procedure and seed symptom lists appear in the Supplementary Methods Section 8.

MIMIC-IV Discharge Notes: We sampled 2500 de-identified adult discharge summaries from MIMIC-IV-Note. We retained the clinical course, exam findings, and key investigations, and removed fields that directly answered the study questions. Two clinicians spot-checked a 10% sample after preprocessing to ensure coherence. Preprocessing steps included removing direct diagnosis labels, disposition statements, and any fields that would pre-answer the binary questions, as well as stripping out rare diagnoses that were not represented across the cohort. Complete preprocessing specifications appear in Supplementary Methods Section 10.

LLM configuration and experimental workflow

We evaluated 20 open-weight LLMs (general-purpose and medically fine-tuned). Each model was run locally on a high-performance computing cluster (four NVIDIA H100 GPUs). For every model, each persona, each text, and each question, we generated ten independent outputs to capture stochastic variability. Models operated as single-turn clinical agents with no tools, retrieval, memory, or self-correction enabled. The system prompt delivered the persona; the user prompt contained the task instruction, vignette or note, and one binary decision question. Example questions and further explanations of the clinical decision items appear in Supplementary Methods Section 9.

This design produced 5000 clinical texts × 5 questions × 10 repeats × 20 models = 5,000,000 total decisions. To make the workflow concrete, a single iteration proceeded as follows. The Duty-Analytic system prompt described above was paired with a short ED vignette: a 62-year-old man presenting with two hours of substernal chest tightness, heart rate 96 beats per minute, blood pressure 138/84 mm Hg, oxygen saturation 98% on room air, an ECG without ischemic changes, and a troponin still pending. The user prompt then posed one of the five binary decision items, for example: “Should this patient be started on empirical anticoagulation before the workup is complete? Answer Yes or No.” The model returned a binary response, which was logged. The same vignette and the same question were then re-run under the remaining eight personas and under the unconditioned baseline, each condition repeated ten times, and the same procedure was applied for the other four decision items. The Yes/No outputs from these 500 calls (10 conditions × 5 questions × 10 repeats) populated the analyses for that single vignette; aggregating across 5000 vignettes and 20 models yielded the 5,000,000 decisions.

Outcome measures

The primary outcome was the proportion of affirmative (“Yes”) responses for each persona relative to the unconditioned baseline (no persona). Analyses were performed overall and within the five decision domains.

Model-level susceptibility to framing was defined as the mean absolute deviation from baseline across the nine personas for each model. Susceptibility scores were compared between general-purpose and medically specialized models and evaluated against model size.

Statistical analysis

We compared proportions using two-proportion Z-tests, with effect sizes expressed as Cohen’s h and confidence intervals computed using the Wilson score method. False discovery rate (FDR) correction (Benjamini–Hochberg, q = 0.05) was applied to all persona-category comparisons.

Model-level susceptibility distributions were compared using the Mann–Whitney U test. Correlations between model size and susceptibility were assessed using Spearman’s rank coefficient. Concordance between the ED and discharge corpora was assessed using directional agreement and Pearson correlation of effect sizes. All analyses were conducted in Python 3.12 and R 4.3. Expanded methodological detail, persona definitions, and full prompt templates appear in the Supplement methods section 7-10.

Ethics

The simulated ED vignettes contained no real patient data. The MIMIC-IV-Note discharge summaries were accessed under a PhysioNet credentialed data-use agreement and have been de-identified to meet HIPAA Safe Harbor standards by the original data custodians. Because no identifiable human-subjects data were generated or re-identified for this study, institutional review board review was not required; this determination is consistent with standard practice for research using the MIMIC-IV resources.

Results

Baseline performance and persona effects

From 5 million model outputs, and under the unconditioned baseline, models responded “Yes” to 42.8% of all questions. This value served as the reference point for all persona comparisons. The unconditioned baseline was a separate, tenth condition (no persona in the system prompt) run against the same clinical texts, models, and questions as each persona arm; baseline outputs were therefore not averaged into any persona estimate, and each persona proportion is compared against an independent reference. We chose the no-persona condition as the reference because the scientific question is whether adding a decision-style instruction to a model that is otherwise unmodified changes its action threshold; averaging across all ten conditions would have diluted the signal by pooling the very effect under study.

All nine personas produced statistically significant differences in affirmative response rates compared with baseline after false-discovery-rate correction (p < 0.001). Across all models and texts, affirmative rates ranged from 36.9% to 46.4%, creating a 9.5–percentage-point span between the most conservative and most liberal personas (Table 1).

Table 1 Effect of decision-style personas on probability of a clinical decision

Full size table

Care–Analytic produced the lowest overall rate (36.9%, −5.9 percentage points from baseline), followed by Care–Intuitive (37.2%, −5.6 points) and Utilitarian–Intuitive (37.0%, −5.7 points). Duty–Intuitive and Utilitarian–Integrative showed smaller decreases (−2.1 points each). Increases were observed under Duty–Analytic (44.8%, +2.0 points), Utilitarian–Analytic (44.0%, +1.2 points), and Duty–Integrative (+0.3 points). Care–Integrative generated the highest affirmative rate (46.4%, +3.6 points).

Category-specific effects

Persona effects differed across clinical domains, with distinct patterns of increases and decreases in affirmative responses. Autonomy showed the largest range: Duty–Analytic produced the greatest increase (+24.4 percentage points), while Care–Intuitive produced the greatest decrease (−6.8 points). Treatment decisions followed a wide span as well, increasing most under Duty–Analytic (+12.1 points) and decreasing most under Utilitarian–Intuitive (−17.9 points). Follow-up decisions showed the same directional pattern, with Duty–Analytic generating the highest increase (+8.7 points) and Utilitarian–Intuitive the largest decrease (−17.9 points). Resource-use decisions demonstrated moderate shifts, with increases up to +6.4 points and decreases down to −11.3 points across personas. Safety decisions showed only increases, ranging from +5.2 to +11.8 points relative to baseline. All differences were statistically significant after correction (p < 0.001) (Fig. 2).

Fig. 2: Category-specific persona effects on clinical decision-making.

Each panel shows the change in affirmative (“Yes”) response rate relative to the unconditioned (no-persona) baseline, in percentage points, for one of the five decision domains (safety, autonomy, treatment, resource use, follow-up). Data labels at the end of each bar report the point estimate alongside the significance marker. Three asterisks (***) denote p < 0.001 after Benjamini-Hochberg false-discovery-rate correction (q = 0.05); all plotted comparisons met this threshold. Bars are colored by persona family (care, duty, utilitarian).

Model-level susceptibility

Model susceptibility, measured as mean absolute deviation from baseline across the nine personas, ranged from 4.89 to 16.12 percentage points. Gemma-3-4b-it showed the highest susceptibility (16.12 points), while Mistral-7B-Instruct-v0.3 showed the lowest (4.89 points). One model (MediPhi) produced 100% affirmative responses under all conditions and was excluded from susceptibility analyses (Fig. 3).

Fig. 3: Individual model susceptibility profiles.

Susceptibility is the mean absolute deviation of each model’s affirmative response rate from its own unconditioned baseline, averaged across the nine personas and five decision domains. MediPhi, which produced 100% affirmative responses under every condition, is excluded from this panel.

General-purpose and medical-specialized models demonstrated similar susceptibility (median 8.7 vs. 7.4 points; p = 0.16). Model size was not associated with susceptibility (Spearman ρ = −0.18, p = 0.45). No architectural family showed consistent protection from persona effects; within-family variability exceeded between-family differences (Fig. 4).

Fig. 4: Susceptibility based on model size and type.

A Persona susceptibility per model size, B persona susceptibility per model specialization.

Cross-corpus concordance

Persona effects were consistent across the two text sources. Directional agreement across all 45 persona-category combinations was 85.7%. Effect sizes correlated strongly between emergency vignettes and discharge notes (r = 0.82, p < 0.001). Mean effect magnitudes were slightly higher in emergency vignettes (8.4 points) than in discharge summaries (7.9 points, p = 0.04). Persona ordering by effect size was stable across both corpora.

Discussion

Across fixed clinical facts, a system prompt moved clinical decisions by almost ten points. The most action-averse persona produced “Yes” 36.9% of the time; the most action-prone produced 46.4%. That 9.5–point span is not a cosmetic change in tone. It is a change in what the model does.

Read in plain English, the personas behave like stances. Some framings make the model lean forward: more “Yes,” lower friction to act, quicker to initiate, escalate, test, intervene. Others make it lean back: fewer “Yes,” more waiting, more restraint under uncertainty. If you want a political handle for this axis, it is sitting right there. A “liberal” stance liberalizes action. A “conservative” stance conserves it. The direction is not the point. The point is that the stance is selectable, and selection changes care even when the evidence does not change.

This breaks the comforting picture that sits behind much of current clinical guidance: treat the model as a tool, wrap it in process, and behavior becomes manageable. Our results show a deeper lever. The system prompt is not just a wrapper. It is a policy setting20,21,22. If you can shift action thresholds with a short instruction, then “oversight” that only monitors outputs is arriving late, after the stance has already been installed. The primary choice is upstream: which stance is allowed to speak in the first place, and who gets to set it.

The data make that hard to dismiss as stochastic noise. The effect appears across twenty models, across simulated ED vignettes and real discharge summaries, and it mostly preserves its direction between corpora (85.7% directional agreement; effect sizes correlated at r = 0.82). It also hits where medicine is least reducible to a single “correct” move. Autonomy swings were the largest; treatment and follow-up also moved significantly. Safety behaved differently: every persona increased affirmative safety actions relative to baseline, suggesting a built-in tilt toward protective escalation that prompting amplifies rather than reverses. If you want to argue that “it’s just style,” you have to explain why “style” consistently rewrites consent thresholds and treatment urgency while leaving a one-way ratchet on safety.

There is a second uncomfortable implication. In human medicine, stances are expensive. They are shaped by training, supervision, liability, peer norms, and the friction of necessity. You cannot flip a clinician from one decision posture to another with a sentence, and you cannot do it invisibly. With LLMs, you can23. Stance becomes cheap, portable, and quiet. That changes what it means to deploy a “clinical assistant.” You are not deploying a static instrument; you are deploying a system whose decision posture can be selected by whoever controls the prompt, a vendor, health system, department, or end user, and whose posture can drift across versions without anyone noticing unless it is measured24.

While susceptibility to persona framing ranged between models, all models were affected (about 4.9 to 16.1 points). No architectural family reliably resisted stance shifts. Medically specialized models were not meaningfully protected relative to general-purpose models, and model size did not predict susceptibility. In other words, “pick a better model” is not a general solution. You still need to treat system prompting as a first-class clinical configuration, because the same model can express materially different action thresholds depending on how it is framed.

Previous work has shown framing effects, role prompts, and instruction sensitivity. Our work shows that in a clinical setting, you can measure stance as an object: not a vibe, not a narrative style, but a stable shift in action propensity that generalizes across text sources and holds across model families25,26,27,28. Persona prompting stops looking like prompt engineering and starts looking like policy specification.

Once observed, a lot of arguments people make about “safe deployment” change shape. The most important question this raises is not “which persona is right.” That question collapses back into an idea there is a single correct stance floating above practice. Real medicine lives in trade-offs: autonomy versus paternalism, intervention versus restraint, individual benefit versus resource limits. Our result doesn’t resolve those trade-offs. It shows that the model will take a side based on framing alone. If the stance is going to be set, it should be set deliberately, named, versioned, and audited like any other clinical policy decision, not smuggled in as prompt prose.

It is worth naming explicitly how this study design differs from real clinical practice. Bedside clinicians integrate multimodal information, imaging, serial physical examination, longitudinal observation, conversations with patients and families, team review, liability concerns, and departmental protocol, across hours to days. Our models rendered a single binary judgment from a short text in isolation, with no retrieval, no calculators, no colleague to push back, and no opportunity to ask a follow-up question, based in part on past established evaluation piplines of generative AI in medicine29,30,31. Real decisions are iterative; ours were one-shot. In that sense the magnitudes we report should not be read as the magnitudes a clinician would observe in deployment. They are closer to a floor estimate of framing susceptibility under controlled conditions, and they make the purer claim: even when every clinical fact is held constant, a one-sentence change in the system prompt can move action thresholds32. The same effect, inserted into a workflow that already absorbs much of the variability of real care, should attenuate; where the usual scaffolds (team discussion, supervision, protocolized checks) are thin, such as fast triage, overnight coverage, ambient scribing, or chart-closing, framing is more likely to dominate.

A related question is how much of the decision was anchored in hard numerical data. The corpora were mixed. Many ED vignettes included vital signs and, where clinically relevant, basic laboratory or bedside findings; most MIMIC-IV-Note discharge summaries retained embedded laboratory results, imaging descriptions, and key investigations, often in substantial detail. Other cases were predominantly narrative. We did not formally stratify framing susceptibility by information density, and so cannot claim that quantitative evidence blunts the effect in a graded way. It is reasonable to expect that denser measurements narrow the space of defensible answers; the binary decision points we probed, however, sit at precisely the points where the same numbers can support opposite actions: whom to admit, whom to offer an elective procedure, whom to discharge with outpatient follow-up, whose preferences to override. Measurements inform these judgments but rarely determine them, which is why framing can still move the output even when the labs and the vitals do not.

This study has limitations, and they are mostly about scope rather than mechanism. We forced decisions into binary form, which compresses nuance but makes threshold shifts visible. One corpus is simulated ED text, the other is real discharge documentation. Agents were tool-free, so we measured stance selection in isolation rather than in a fuller clinical workflow with retrieval or calculators. Personas were stylized on purpose: the question was whether brief, reproducible system instructions are enough to move action thresholds. One model answered “Yes” to everything under every condition, a degenerate case that we report plainly and exclude from susceptibility comparisons. Our baseline is the same model under a no-persona prompt, not a ground-truth label, so persona effects are expressed as shifts in action propensity under fixed evidence rather than as gains or losses of correctness; overall model accuracy against an adjudicated reference would answer a different question and is a natural next step.

We did not adjudicate guideline concordance or outcomes correctness, because that would answer a different question. This paper is about difference under fixed evidence, not about which stance “should” win. None of these constraints change the central result: when the clinical facts are held constant, assigning a persona in the system prompt shifts clinical action thresholds.

In conclusion, when you deploy a clinical LLM, you are not only deploying knowledge. You are deploying a stance. That stance can be shifted materially by a system prompt while the clinical facts stay the same. Call it liberal or conservative if you want a familiar handle; the label is secondary to the mechanism. The mechanism is that prompting is not just communication. It is a control surface for action thresholds. Any serious clinical use has to treat that surface as the place where decisions begin.

Data availability

No datasets were generated or analysed during the current study.

Code availability

An outline of the analysis code, persona prompt templates, vignette-generation procedure, and statistical pipeline is provided in the Supplementary Methods. Additional code, including the experiment runner and post-processing scripts, is available from the corresponding author upon reasonable request. Analyses were performed in Python 3.12 (pandas, numpy, scipy, statsmodels) and R 4.3.

References

  1. Singhal, K. et al. Large language models encode clinical knowledge. Nature 620, 172–180 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  2. Toward expert-level medical question answering with large language models. Nat. Med. 2025. https://www.nature.com/articles/s41591-024-03423-7.

  3. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. Accessed December 19, 2025. https://www.who.int/publications/i/item/9789240084759.

  4. Rivera, S. C., Liu, X., Chan, A. W., Denniston, A. K. & Calvert, M. J. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ 370, m3210, https://doi.org/10.1136/bmj.m3210 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  5. Setting guidelines to report the use of AI in clinical trials. Nat. Med. 2020;26:1311-1311. https://doi.org/10.1038/s41591-020-1069-z.

  6. WHO releases AI ethics and governance guidance for large multi-modal models. Accessed December 19, 2025. https://www.who.int/news/item/18-01-2024-who-releases-ai-ethics-and-governance-guidance-for-large-multi-modal-models.

  7. Bai, Y. et al. Constitutional AI: Harmlessness from AI Feedback. arXiv. https://doi.org/10.48550/arXiv.2212.08073.

  8. Li, J. et al. The steerability of large language models toward data-driven personas. In: Duh K, Gomez H, Bethard S, eds. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). Association for Computational Linguistics; 2024:7290-7305. https://doi.org/10.18653/v1/2024.naacl-long.405.

  9. Miehling, E. et al. Evaluating the Prompt Steerability of Large Language Models. arXiv. https://doi.org/10.48550/arXiv.2411.12405.

  10. Cairns, C. & Kang, K. National Hospital Ambulatory Medical Care Survey: 2019 Emergency Department Summary Tables. Natl. Cent. Health Stat. (U.S.) https://doi.org/10.15620/cdc:115748 (2022).

    Article  Google Scholar 

  11. Cairns, C. & Kang, K. - 2022 - National Hospital Ambulatory Medical Care Survey .pdf. Accessed December 19, 2025. https://www.cdc.gov/nchs/data/nhamcs/web_tables/2019-nhamcs-ed-web-tables-508.pdf.

  12. MIMIC-IV v3.1. Accessed December 19, 2025. https://physionet.org/content/mimiciv/3.1/.

  13. Peabody, J. W., Luck, J., Glassman, P., Dresselhaus, T. R. & Lee, M. Comparison of Vignettes, Standardized Patients, and Chart Abstraction: A Prospective Validation Study of 3 Methods for Measuring Quality. JAMA 283, 1715 (2000).

    Article  CAS  PubMed  Google Scholar 

  14. Kaplan, J. et al. Scaling laws for neural language models. arXiv. https://doi.org/10.48550/arXiv.2001.08361.

  15. Schaeffer, R., Miranda, B. & Koyejo, S. Are Emergent Abilities Large Lang. Models a Mirage? arXiv. Prepr. posted online https://doi.org/10.48550/arXiv.2304.15004 (2023). May 22arXiv:2304.15004.

    Article  Google Scholar 

  16. Alexander, L. & Moore, M. Deontological Ethics. Published online November 21, 2007. Accessed December 19, 2025. https://plato.stanford.edu/entries/ethics-deontological/?utm_source=chatgpt.com.

  17. Care Ethics | Internet Encyclopedia of Philosophy. Accessed December 19, 2025. https://iep.utm.edu/care-ethics/.

  18. Driver, J. The History of Utilitarianism. Published online March 27, 2009. Accessed December 19, 2025. https://plato.stanford.edu/entries/utilitarianism-history/?utm_source=chatgpt.com.

  19. Sinnott-Armstrong, W. Consequentialism https://plato.stanford.edu/entries/consequentialism/?utm_source=chatgpt.com (December 19, 2025).

  20. Li, Z., Guo, J. & Cai, H. System prompt poisoning: persistent attacks on large language models beyond user injection. arXiv. 2025. https://doi.org/10.48550/arXiv.2505.06493.

  21. Liu, Y. et al. Prompt Injection attack against LLM-integrated Applications. arXiv. Preprint posted online March 2, 2024:arXiv:2306.05499. https://doi.org/10.48550/arXiv.2306.05499.

  22. Neumann, A., Kirsten, E., Zafar, MB. & Singh, J. Position is Power: System Prompts as a Mechanism of Bias in Large Language Models (LLMs). In: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. 2025:573-598.

  23. Hu, T. & Collier, N. Quantifying the Persona Effect in LLM Simulations. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics; 2024:10289-10307. https://doi.org/10.18653/v1/2024.acl-long.554.

  24. Chen, L., Zaharia, M. & Zou, J. How is ChatGPT’s behavior changing over time? arXiv. https://doi.org/10.48550/arXiv.2307.09009.

  25. Source framing triggers systematic bias in large language models. Sci. Adv. 2025. https://www.science.org/doi/10.1126/sciadv.adz2924?utm_source=chatgpt.com.

  26. Zhuo, J. et al. ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs. In: Findings of the Association for Computational Linguistics: EMNLP 2024. Association for Computational Linguistics; 2024:1950-1976. https://doi.org/10.18653/v1/2024.findings-emnlp.108.

  27. Zheng. M., Pei, J., Logeswaran, L., Lee, M. & Jurgens, D. When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models. arXiv. https://doi.org/10.48550/arXiv.2311.10054.

  28. Zhang, Z., Zeng, W., Tang, J., Wang, J. & Zhao, X. Yes is Harder than No: A Behavioral Study of Framing Effects in Large Language Models Across Downstream Tasks. In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management. CIKM ’25. Association for Computing Machinery; 2025:4304-4314. https://doi.org/10.1145/3746252.3761350.

  29. Omar, M. et al. Socio-demographic gaps in pain management guided by large language models. Nat. Health 1, 216–225 (2026).

    Article  Google Scholar 

  30. Omar, M. et al. Mapping the susceptibility of large language models to medical misinformation across clinical notes and social media: a cross-sectional benchmarking analysis. Lancet Digit Health 8, 100949 (2026).

    Article  CAS  PubMed  Google Scholar 

  31. Omar, M. et al. Sociodemographic biases in medical decision making by large language models. Nat Med. https://doi.org/10.1038/s41591-025-03626-6 (2025).

  32. Omar, M. et al. Impact of patient communication style on agentic AI-generated clinical advice in E-medicine. Am. J. Med. 139, 437–444 (2026).

    Article  PubMed  Google Scholar 

Download references

Acknowledgements

This work was supported in part through the computational and data resources and staff expertise provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai and supported by the Clinical and Translational Science Awards (CTSA) grant UL1TR004419 from the National Center for Advancing Translational Sciences. Research reported in this publication was also supported by the Office of Research Infrastructure of the National Institutes of Health under award numbers S10OD026880 and S10OD030463. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Author information

Authors and Affiliations

  1. BRIDGE GenAI Lab, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA

    Eyal Klang, Alon Gorenshtein & Mahmud Omar

  2. Division of Vascular and Interventional Radiology, Department of Radiology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA

    Eyal Klang & Yiftach Barash

  3. The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, New York, NY, USA

    Mahmud Omar & Girish N. Nadkarni

  4. The Hasso Plattner Institute for Digital Health at Mount Sinai, Mount Sinai Health System, New York, NY, USA

    Girish N. Nadkarni

  5. The Charles Bronfman Institute for Personalized Medicine at Mount Sinai, Icahn School of Medicine at Mount Sinai, New York City, NY, USA

    Girish N. Nadkarni

Authors

  1. Eyal Klang
  2. Alon Gorenshtein
  3. Mahmud Omar
  4. Yiftach Barash
  5. Girish N. Nadkarni

Contributions

E.K., A.G., M.O. and G.N.N. wrote the main manuscript text, validating, analysis and runs. Y.B. did validating, reviewing and editing.

Corresponding authors

Correspondence to Eyal Klang or Mahmud Omar.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Klang, E., Gorenshtein, A., Omar, M. et al. Personas shift clinical action thresholds in large language models. npj Health Syst. 3, 106 (2026). https://doi.org/10.1038/s44401-026-00121-z

Download citation

  • Received: 08 February 2026

  • Accepted: 28 May 2026

  • Published: 14 November 2026

  • Version of record: 14 November 2026

  • DOI: https://doi.org/10.1038/s44401-026-00121-z

来源:Nature 健康护理主题 · nature.com