Does “You are an expert” lower AI accuracy? The pitfalls of prompt design revealed by USC research

As “prompt engineering” for mastering AI gains attention, one widely adopted technique is role setting (persona instructions) like “You are an expert in 〇〇.”

However, the latest research reveals that this approach can actually reduce AI accuracy in certain cases.

This counterintuitive result hides important mechanisms related to AI’s learning structure. This article explains the research findings and practical approaches to “correct prompt design” for real-world use.

The pitfalls of “expert personas” revealed by USC research

In 2026, a research team at the University of Southern California (USC) published a preprint paper titled “Expert Personas Improve LLM Alignment but Damage Accuracy.”

The study compared benchmark performance across 6 AI models, including Llama-3.1-8B and Qwen2.5-7B, by applying various persona prompts.

Shocking results: Expert settings lowered scores

According to the paper, in tasks requiring accuracy such as mathematics and coding, the correct answer rate with expert personas (approximately 68.0%) was lower than the baseline with no persona setting (approximately 71.6%).

This is the opposite of what many users expect: “telling it’s an expert should yield higher accuracy.”

“Persona prefixes switch the model’s processing from fact retrieval mode to instruction-following mode”
— Zizhao Hu, PhD candidate at USC (lead author of the paper)

Why does the “expert” instruction backfire?

To understand this phenomenon, we need to grasp that AI acquires different abilities through two-stage training.

AI capabilities have a “two-layer structure”

Layer 1: Abilities acquired through pre-training

• Mathematical reasoning and logical thinking
• Programming and coding
• Factual knowledge and encyclopedic information

These are abilities acquired during the “pre-training” phase where the model learns from massive text data.

Layer 2: Abilities acquired through fine-tuning

• Style and tone adjustment
• Instruction following
• Safety filtering (refusal of harmful content)

These are abilities adjusted later through techniques like “RLHF (Reinforcement Learning from Human Feedback)” using human feedback.

Persona instructions press the “wrong switch”

The instruction “You are an expert” strongly activates the instruction-following mode of Layer 2.

As a result, instead of drawing on factual knowledge stored in Layer 1, the model focuses on generating text that seems appropriate.

You cannot add “expert knowledge” through prompts. It’s impossible to add knowledge the model doesn’t have, and instead, it hinders access to the knowledge the model actually possesses.

Tasks where persona instructions “work” and “don’t work”

Importantly, expert personas aren’t bad in all situations. The research also shows that effects vary significantly by task type.

❌ Tasks where persona instructions are counterproductive

These are “pre-training dependent” tasks.

• Solving math problems
• Writing and debugging code
• Fact-checking
• Answering logic puzzles

✅ Tasks where persona instructions are effective

These are “alignment dependent” tasks.

• Writing text in specific tones or styles (e.g., “professional email”)
• Role-playing and simulation
• Data structuring and formatting
• Controlling safety and ethical responses

Additionally, JailbreakBench (an evaluation metric for ethical content filters) confirmed that using expert personas significantly improves model safety scores.

So, how should you write? “The right way to give expertise”

Based on the research findings, the goal in prompts should be to give “behavioral standards” rather than “titles.”

Before (low-effectiveness prompt)

You are an experienced lawyer. Provide perfect legal advice.

This prompt gives a “role” but doesn’t specify what and how to output. The model focuses on acting “lawyer-like” and tends to prioritize “appropriate style” over factual accuracy.

After (high-effectiveness prompt)

When answering the following question, please follow these criteria:
- Use legal terminology accurately
- If definitions are ambiguous, explicitly state so
- If you cannot be certain, note "it is possible that..."
- If there are laws or articles to reference, specify them concretely

The key is changing from the “expert” title to specific behavioral instructions that ensure output quality.

Practical rewriting patterns

Practical application: Decision flowchart

When designing prompts, ask yourself the following questions.

① Does the task require “accuracy” or “style”?

• Accuracy needed (math, code, fact-checking) → Avoid persona instructions. Write specific output criteria.
• Style needed (writing style, tone, structure) → Persona instructions are effective. But don’t apply excessive pressure.

② Are expressions like “perfectly as an expert” included?

Excessive pressure words like “perfectly” and “as a professional” tend to lead the model into “acting mode,” so it’s recommended to avoid them.

Summary: Give “standards” rather than “titles”

To summarize the USC research points:

• The instruction “You are an expert” can lower accuracy in tasks like math and coding
• The reason is that persona instructions activate the model’s “instruction-following mode,” hindering fact retrieval capabilities
• On the other hand, it’s effective for “alignment dependent” tasks like style and tone adjustment
• Rather than giving titles, specific behavioral instructions about “what to output and by what standards” are effective

AI output quality can change significantly with just how you write prompts. Let go of the assumption that “expert settings improve quality,” and shift your focus to verbalizing output standards.

That’s the first step to truly mastering AI.

References

• Zizhao Hu et al., “Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM,” USC preprint (2026)
• The Register, “Telling an AI model that it’s an expert makes it worse” (March 2026)