Stop Asking AI to 'Act Like an Expert' - Science Says It Doesn't Work
You've seen the advice everywhere. "You are a world-class marketing strategist..." "Act like an experienced data scientist..." "Pretend you're a senior software engineer..."
OpenAI's prompting guide recommends it. Google's Gemini docs suggest it. Anthropic's Claude examples use it. Every AI tutorial on YouTube starts with "tell the AI what role to play."
Wharton's AI lab just proved it's cargo cult advice.
The Study: 6 Models, Zero Improvement
Researchers at Wharton tested expert personas across six major LLMs - GPT-4o, Claude Opus, Gemini Advanced, and three others. They gave each model graduate-level questions from business, medicine, and law.
Half the prompts used expert personas: "You are a world-class economist with decades of experience..."
Half used plain instructions: "Answer this economics question."
The result? Zero consistent improvement. Sometimes personas made answers worse.
Not marginally worse. Not "depends on the task" worse. Just worse.
The study appears in Wharton's AI research today, led by Professor Ethan Mollick, who's been studying AI adoption since GPT-2.
Why Everyone Got It Wrong
The expert persona pattern came from early GPT-3 days when models were unreliable and needed heavy prompt engineering. Early researchers found that adding context sometimes helped.
AI companies picked it up and put it in their docs. Tutorials copied the pattern. Now it's gospel.
But modern models don't need it. They're trained on enough expert-level text that telling them "you're an expert" adds nothing but tokens.
Worse, personas can introduce confusion. When you tell Claude "you're a senior developer," does that mean write senior-level code or explain things like you're talking to a senior developer? The model has to guess.
What Actually Works
The Wharton study identified three prompting patterns that consistently improve output:
1. Clear, specific instructions
Don't: "You are an expert data analyst. Analyze this."
Do: "Calculate the median, identify outliers above 2 standard deviations, and explain what might cause them."
2. Examples of what you want
Don't: "You're a professional copywriter. Write product descriptions."
Do: "Write a product description like this: [example]. Match the tone, length, and structure."
3. Structured output formats
Don't: "Act like a project manager and give me a timeline."
Do: "Create a timeline in this format: Phase | Duration | Deliverables | Dependencies"
Notice the pattern? You're not telling the AI who to be. You're telling it what to do.
The Real Cost of Bad Prompts
This isn't just academic. If you're using AI daily, bad prompting habits waste time.
Every "act like an expert" persona adds tokens to your prompt. On paid APIs, that's money. On free tiers, that's quota.
More importantly, vague personas force you into more back-and-forth. You ask for "expert analysis," get something generic, then clarify what you actually meant. Three messages instead of one.
The Wharton study tested this. Participants using clear instructions got usable answers on the first try 73% of the time. Participants using expert personas? 52%.
What to Do Right Now
If you've been using expert personas, you can improve your AI results immediately:
Stop doing this:
- "You are a world-class [anything]..."
- "Act like a professional [role]..."
- "Pretend you're an expert in [field]..."
Start doing this:
- Lead with the task: "Write," "Analyze," "Compare," "Summarize"
- Specify the format: bullet points, table, paragraph, JSON
- Give an example of what good looks like
- Include constraints: word count, reading level, required elements
Example transformation:
Before: "You are an experienced email marketer. Write a promotional email for our new product."
After: "Write a 150-word promotional email for [product]. Include: benefit-focused subject line, one customer pain point, one product feature that solves it, clear CTA. Tone: friendly and direct, like Grammarly's emails."
Same request. Half the tokens. Better output.
The Bigger Pattern
This study matters beyond prompting tips. It's evidence that AI is maturing.
Early GPT-3 needed tricks and hacks. Modern models want clarity.
The tools are getting smarter. The prompts should get simpler.
If you're still using 2022 prompting patterns in 2026, you're fighting the model. Stop performing expertise. Start requesting deliverables.
The "act like an expert" era is over. The AI already knows what experts sound like. Your job isn't to assign roles. It's to be specific about what you need.
Try It
Take your most common AI prompt. Remove every "you are" and "act like" phrase. Replace them with clear instructions and constraints.
Run both versions. Compare the outputs.
Wharton's research says the simpler prompt will win.
Source: Wharton Knowledge - "Do Personas Improve AI Performance? New Research Says No" (April 28, 2026) | Professor Ethan Mollick's AI research lab