Being Rude To LLMs Increases Their Performance?!

A 2025 research paper titled 'Mind Your Tone' investigated how prompt politeness affects LLM accuracy. Testing GPT-4o on 50 multiple-choice questions across five tones from very polite to very rude, the study found that ruder prompts consistently yielded higher accuracy, with very rude prompts scoring 84.8% versus 80.8% for very polite ones. The findings come with significant caveats: small dataset, single model, and only multiple-choice format. The post raises broader questions about why tone affects LLM performance and whether rude AI interactions could reinforce rude communication habits in users.