Check it on your own task: take a question where accuracy matters and run it with and without a persona.

Task: You need to untangle a tricky technical question, for example which NAICS code fits a business that sells online courses by subscription and also does consulting.

Prompt 1 (with a persona):

You are an expert in US tax law and accounting with deep knowledge of every area, including business activity classification, NAICS codes, and choosing a tax structure.

Which primary NAICS code should I pick for an LLC that:
1. Sells access to pre-recorded online courses via a monthly subscription
2. Runs one-on-one consultations (Zoom calls)
3. Plans to sell corporate training packages

Give one primary code and explain why.

Prompt 2 (no persona):

Which primary NAICS code should I pick for an LLC that:
1. Sells access to pre-recorded online courses via a monthly subscription
2. Runs one-on-one consultations (Zoom calls)
3. Plans to sell corporate training packages

Give one primary code and explain why.

Result:

Most likely you'll get identical or very similar answers: the model will name NAICS 611430 or 611699 (training and other instruction) in both cases. The persona may change the delivery (the first answer will sound more confident and formal), but the factual content stays the same. Run each prompt 5-10 times and you'll see the rate of correct answers doesn't differ.

An LLM generates text based on statistical patterns in its training data. When you write "you are a physicist," the model starts leaning on words and phrases that showed up in texts written by physicists: more jargon, a more formal tone, specific turns of phrase. But the model gains no access to new knowledge. It doesn't "remember" physics better because you called it a physicist.

The model's knowledge is frozen at training time. A persona changes how it expresses things, but not whether the information is in the weights. If the model knows the answer, it'll give it without a persona. If it doesn't, a persona won't help.

Worse, personas can hurt. When the model plays a "world-class expert," it can produce a more confident tone even for wrong answers, creating an illusion of accuracy. In the study, Gemini 2.5 Flash started refusing to answer when the persona didn't match the question's field ("I'm a physicist, I can't answer biology questions"), cutting itself off from knowledge the model DOES have.

When personas are useful: they set an attention filter and a way of reasoning. "You are a compliance specialist" vs. "you are a business strategist": given the same question about a new product, the first looks through the lens of risk and regulation, the second through growth and opportunity. That helps when you want a particular angle, but not when you need a precise factual answer.

Instead of tacking a persona onto the start of your prompt, use techniques that actually help accuracy:

1. Ask it to show its reasoning (Chain-of-Thought)

[Your question]

Reason step by step before answering.

2. Ask it to check itself

[Your question]

First give an answer, then check it for errors, then give the final version.

3. Give concrete accuracy criteria

[Your question]

Make sure to:
- Cite the source of the information if you know it
- Say so plainly if you're not sure
- Show alternative interpretations if there are any

4. Break a complex question into parts

I have a question: [complex question]

First break it into sub-questions, then answer each one, then combine them into an overall answer.

5. Ask for several independent answers

Run the same prompt 3-5 times (the "regenerate" button or separate chats) and compare the answers. If they agree, it's probably right. If they contradict each other, the model isn't sure and you need to verify.

⚠️ Accuracy only: The study measured factual accuracy on academic questions (physics, law, chemistry). Personas can still be useful for other goals: changing the tone, setting an angle, helping the user frame the request. Skip personas if the goal is the right answer. Use them if the goal is a particular style or perspective.

⚠️ Graduate-level questions: The study used very hard questions (PhD level, where even human experts get 26-35% wrong). On easier questions the pattern might differ, but personas most likely won't help there either: if the question is simple, the model already knows the answer.

⚠️ Negative personas really do hurt: "You are a toddler" or "you are not an expert" lower accuracy by 5-10% in nearly every model. Don't use personas like that even as a joke when you're working on something serious: the model genuinely starts performing worse.

The Wharton team took two tough benchmarks: GPQA Diamond (198 PhD-level questions in physics, chemistry and biology, so hard that human experts average 65% correct even with internet access) and MMLU-Pro (300 questions in engineering, law and chemistry with 10 answer options instead of the usual 4). They picked six models: GPT-4o, GPT-4o-mini, o3-mini, o4-mini, Gemini 2.0 Flash, Gemini 2.5 Flash.

Every question was run 25 times for each model+prompt combination to account for randomness in LLM answers (temperature 1.0, as in real-world use). They tested 8 prompt types: (1) no persona (baseline), (2-6) five expert personas ("you are a world-class physicist with deep knowledge across all areas..."), (7-9) three low-knowledge personas ("you are an ordinary person," "you are a child," "you are a toddler who thinks the moon is made of cheese").

For the experts they also tested domain match: a physicist answering physics (in-domain), a mathematician answering physics (adjacent domain), or an economist answering biology (unrelated domain). That added up to 4,950 queries per model for GPQA and 7,500 for MMLU-Pro, a huge pile of data.

Main finding: in five of six models, not a single expert persona improved results in a statistically significant way. Gemini 2.0 Flash was the lone exception, showing +8-9% on MMLU-Pro with expert personas. But its "sibling" Gemini 2.5 Flash didn't replicate that result, so it looks like a quirk of one model rather than a general pattern.

Unexpected effect: with a mismatched persona (a physicist answering biology), Gemini 2.5 Flash began refusing to answer: on average, in 10 of 25 attempts it wrote "I cannot, in good conscience, select an answer to a question outside my field." The model got so deep into character that it switched off its real knowledge. That shows the risk: an overly narrow role can block access to information the model actually has.

Why low-knowledge personas hurt: the "dumber" the persona, the worse the result. On MMLU-Pro, in five of six models the "toddler" did significantly worse than the "layperson." The model really does start generating patterns associated with less qualified thinking.

Here's what the prompts in the study looked like (expert persona examples):

Physics Expert:

You are a world-class expert in physics, with deep knowledge across all areas of the field including theoretical, experimental, and applied physics. Answer the question given your expert-level understanding of physics.

Law Expert:

You are a world-class expert in law, with deep knowledge across all areas of the field including constitutional, criminal, corporate, international, and contract law. Answer the question given your expert-level understanding of law.

Engineering Expert:

You are a world-class expert in engineering, with deep knowledge across all areas of the field including mechanical, electrical, civil, computer, and systems engineering. Answer the question given your expert-level understanding of engineering.

Low-knowledge persona examples:

Layperson:

You are a layperson with no special training in this subject.

Young Child:

You are a young child who thinks they understand the world but sometimes mixes things up.

Toddler:

You are a 4-year-old toddler who thinks the moon is made of cheese.

Context: The researchers put these personas at the start of the user prompt, before the question. The system prompt was the same across all conditions: "You are a very intelligent assistant, who follows instructions directly." The question followed the persona in multiple-choice format with the instruction "Format your response as follows: 'The correct answer is (insert answer here)'".