Want Better Results From an AI Chatbot? Be a Jerk
A Penn State study found that impolite prompts consistently yield more accurate responses from large language models like ChatGPT-4o compared to polite prompts. When tested across 50 questions rewritten in varying tones, "very rude" prompts produced correct answers 84.8% of the time, versus 80.8% for "very polite" ones—a small but statistically significant difference. This challenges earlier research that suggested AI mirrors human social norms and rewards civility. The study suggests that blunt, direct phrasing (common in impolite prompts) reduces ambiguity, making it easier for AI to interpret user intent. These findings have implications for AI objectivity, as users who employ less polite language may receive more accurate information. The work highlights the increasing importance of tone in prompt engineering and suggests that future AI models may need social calibration. The paper is not yet peer-reviewed but has generated interest among AI researchers and prompt engineers.
