Prompt Politeness Affects LLM Accuracy

A study investigating how prompt politeness affects large language model accuracy found that rude and very rude prompts outperformed polite ones, with accuracy ranging from 80.8% for very polite prompts to 84.8% for very rude prompts on multiple-choice questions across mathematics, science, and history. This counterintuitive result suggests that newer LLMs like ChatGPT 4o may respond differently to tonal variation than previously observed in earlier studies. The findings highlight the importance of studying pragmatic aspects of prompting and raise questions about the social dimensions of human-AI interaction.