Politeness Penalty: Research Shows Rude Prompts Outperform Polite Ones on ChatGPT 4o

Summary

A new academic study challenges conventional wisdom about how users should interact with large language models. Researchers tested how prompt politeness affects ChatGPT 4o's accuracy on multiple-choice questions spanning mathematics, science, and history. The results were surprising: rude and very rude prompts consistently outperformed polite and very polite ones, with accuracy improving from 80.8% for very polite prompts to 84.8% for very rude prompts.

The study analyzed 250 prompts—50 base questions rewritten into five tone variants (Very Polite, Polite, Neutral, Rude, and Very Rude)—and applied paired sample t-tests to validate statistical significance. The findings contradict earlier research suggesting that rudeness produces poorer outcomes, indicating that newer LLM architectures may process tonal variation differently than previous models. The research highlights an important but underexplored aspect of prompt engineering: the pragmatic and social dimensions of human-AI interaction.

• Raises questions about how social dimensions and tone influence model behavior

Editorial Opinion

This counterintuitive result suggests that LLM designers and users should reconsider assumptions about what constitutes 'human-friendly' interaction design. While the 4% accuracy improvement is modest, it fundamentally challenges the notion that politeness universally improves AI responses and invites deeper investigation into how models process pragmatic language features. Future research should explore whether this pattern holds across different model architectures and whether it reflects genuine differences in how newer models interpret social cues.