Key answer
Expert roles sometimes improve performance, but they are not a generally reliable way to improve accuracy. Studies have found improvements, no significant change and declines. Neither “it works” nor “it only changes tone” adequately describes the evidence.
Full explanation
Consider experiments that directly studied this question:
| Study | Experimental scope | Main finding |
|---|---|---|
| Findings of EMNLP 2024: *When “A Helpful Assistant” Is Not Really Helpful* | 162 roles, four model families and 2,410 factual questions | Adding roles did not consistently improve overall performance. Some roles helped some questions, but predicting which would work was difficult. Paper |
| EMNLP 2025: *Principled Personas* | Nine open-weight models and 27 knowledge and reasoning tasks | Expert roles improved performance or made no significant difference on most tasks, with some declines; narrower expertise was not consistently better. Paper |
| Wharton Generative AI Labs technical report, 2025 | Six OpenAI/Google models; 198 GPQA Diamond and 300 MMLU-Pro questions; 25 repetitions per question for each model/prompt condition | Even domain-matched roles produced no accuracy benefit consistent across models. These were difficult objective questions; results cannot directly be generalized to writing, teaching or simulated conversations. Report |
Why do some papers find a benefit? What they actually added matters.
NAACL 2024's *Better Zero-Shot Reasoning with Role-Play Prompting* reported improved reasoning. Its method combined a task-relevant role description with a selected role-confirmation response generated by the model, using this interaction as problem-solving context. Its main experiments used gpt-3.5-turbo-0613.
This supports the effectiveness of that role-prompting method in those experiments, rather than the claim that adding “you are an expert” to any question improves accuracy. The role, task description and constructed interaction changed together. Methods
Conversely, improvements cannot all be dismissed as merely sounding more expert. *Principled Personas* measured verifiably correct answers and found real accuracy gains for some model/task combinations. The issue is that benefits are inconsistent, and more specific or prestigious professional titles are not necessarily better. Results
There is no established unified mechanism explaining role effects. “It activates expert knowledge inside the model,” for example, is an interpretation rather than a confirmed fact.
The 2026 PRISM preprint further found that expert roles could improve some writing, formatting and intent-following results while harming knowledge multiple-choice performance in its tests. The authors propose explanations involving training and task type, but these do not establish one mechanism across all models. Higher evaluator scores for generated content also do not necessarily mean greater factual accuracy. Study
What should you do in practice?
You do not need an expert role before every question. If your goal is accuracy, current research does not support “you are an expert in…” as a reliably effective preliminary step. It may help, do little or make results worse. If you need a role's working behavior, such as a mock interviewer or an editor's review, a role can express that requirement. Still check role-following and correctness separately.
For example, these prompts serve different purposes:
· “You are a senior interviewer; tell me whether this experience is genuine.” The expert identity is not evidence of a correct judgment.
· “Act as an interviewer and ask detailed follow-up questions about this experience, focusing on my contribution and verifiable details.” The role specifies the interaction, while the rest specifies the work.
This is research-informed teaching advice. It does not mean the second wording has been proven to score higher on every task. A role can be part of a task, but it cannot guarantee accuracy.
References
1. Zheng, M., Pei, J., Logeswaran, L., Lee, M., & Jurgens, D. (2024). *When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models.* Findings of EMNLP 2024, 15126–15154. Paper
2. de Araujo, P. H. L., Röttger, P., Hovy, D., & Roth, B. (2025). *Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance.* EMNLP 2025, 26857–26886. Paper
3. Basil, S., Shapiro, I., Shapiro, D., Mollick, E. R., Mollick, L., & Meincke, L. (2025). *Prompting Science Report 4: Playing Pretend: Expert Personas Don’t Improve Factual Accuracy.* Technical report, arXiv:2512.05858 (not a peer-reviewed conference paper). Report
4. Kong, A., et al. (2024). *Better Zero-Shot Reasoning with Role-Play Prompting.* NAACL 2024. Paper
5. Hu, Z., Rostami, M., & Thomason, J. (2026). *Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM.* arXiv:2603.18507, citing v1 (preprint). Study