Why Telling AI It’s a Genius Might Be Making It Dumber
You’ve probably seen it everywhere — prompting guides that open with “You are an expert full-stack developer…” or “You are a world-class data scientist…” It feels intuitive. You want expert output, so you ask for an expert.
New research suggests this might be working against you.
What the Research Found
A pre-print paper from researchers at USC found that persona-based prompting is task-dependent and that explains the inconsistent results people see in practice.
- Alignment tasks (writing, tone, safety): Expert personas help.
- Factual tasks (math, coding, Q&A): Expert personas hurt.
Using the MMLU benchmark, models under an expert persona underperformed the base model on every subject 68.0% vs. 71.6% accuracy.
The reason? Telling a model it’s an expert doesn’t give it expertise. It just shifts the model into instruction-following mode, which competes with the factual recall it would otherwise use.
The Simple Rule
The researchers put it plainly:
“When you care about alignment — safety, rules, structure — be specific. If you care about accuracy and facts, don’t add anything. Just send the query.”
Skip the expert intro when you need real answers. Save the persona for shaping style, tone, and behavior.