The idea of “please” and “thank you” might feel natural when talking to people, but new research suggests that being rude may actually boost ChatGPT accuracy. In a recent paper, researchers found that impolite or aggressive-toned prompts produced better results on certain tasks using ChatGPT-4o. This article unpacks the findings, context, implications and caveats of this surprising development.
The Study: Tone Matters in Prompt Engineering
A team at Pennsylvania State University (Penn State), led by Om Dobariya and Akhil Kumar, investigated how the tone of prompts affects performance of large language models (LLMs). Their paper “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy” created a test set of 50 base questions (math, science, history) and rewrote each into five tone variants: Very Polite, Polite, Neutral, Rude, Very Rude. That resulted in 250 unique prompts.
They ran these prompts through ChatGPT-4o, measured accuracy (correct answer rate) and found a notable trend:
- Very Polite: ~80.8 % accuracy
- Polite: ~81.4 %
- Neutral: ~82.2 %
- Rude: ~82.8 %
- Very Rude: ~84.8 % accuracy
In short: The more direct and less polite the phrasing, the higher the accuracy—at least in this experimental setting.
Why Could This Be Happening?
The researchers and commentators suggest a few reasons why a ruder tone might help:
- Polite phrasing often adds extra words or indirect phrasing (“Could you please…?”, “Would you be able to…?”) which may dilute clarity. A blunt command may signal intent more directly.
- The model may interpret the more forceful tone as a clearer signal of task urgency or expectation, helping it focus on the question and filter out filler. Digital Trends
- Newer LLMs (like GPT-4o) may have been trained with many variations of prompt styles and may respond better to direct, unambiguous commands than to overly polite or verbose ones.
Important Caveats & Context
While the result is interesting, it’s crucial to recognize limitations:
- The study was limited to multiple-choice questions in specific domains (math, science, history) and used ChatGPT-4o under controlled conditions. Results may differ for creative tasks, open-ended prompts, or other models.
- “Rude” in this study means direct, commanding, or slightly aggressive phrasing—not necessarily highly abusive or hateful language. Extreme toxicity may trigger model decline or refusal.
- The effect size (~4 percentage-point gain) is modest and specific; it doesn’t mean being rude universally guarantees better performance for every task.
- There are ethical and usability implications: encouraging rudeness may impact user behavior and human-AI dynamics in undesirable ways.
Implications for Prompt Engineering & AI Use
For users, developers, and organizations working with AI:
- Prompt clarity still matters: The tone effect underlines that how you ask a question influences output—not just what you ask.
- Direct commands can be more effective: If you need a precise, factual answer, framing your prompt more directly may help.
- Ethics and human-AI interaction count: While the model doesn’t have feelings, promoting rude interactions could shape user habits and norms. Good AI design still values respectful interaction.
- Model and task-specific strategies: Different models, domains and tasks may respond differently to tone. For generative, creative, or multi-turn dialogues, politeness or collaborative phrasing might still be beneficial.
Final Thoughts
The finding that being rude can boost ChatGPT accuracy challenges our assumptions about AI interaction etiquette. It reminds us that large language models operate differently than humans—they don’t care about niceties, they respond to signal clarity. That said, the benefits of rudeness are narrow, the ethical trade-offs real, and politeness still has a strong place in human-AI ecosystems.
For anybody working with AI prompts, the takeaway is: focus on clear, direct instructions. Whether you add “please” is less important than making sure the model sees un-ambiguous intent. But being downright rude? That’s optional—and perhaps better left out of polite company.
