Why being polite to AI might be hurting your results

A new study suggests that being polite to artificial intelligence systems may not be the best approach when trying to get accurate results. Researchers found that the tone used when prompting large language models (LLMs) can significantly affect their accuracy. In a surprising twist, prompts that were less polite or even rude performed better than polite ones. This is a critical finding that challenges a long-held assumption about how humans should interact with AI.

The study, which tested the performance of ChatGPT-4o on multiple-choice questions, found that accuracy increased as prompts became less polite. Very polite prompts achieved an average accuracy of 80.8%, while very rude prompts reached 84.8%, a nearly four-percentage-point improvement. The study suggests that the difference in performance is not due to the tone itself, but rather the way it affects the structure and clarity of the prompt.

Experts say that the findings of this study have significant implications for how we build and interact with AI systems. The study raises questions about how LLMs process language and why certain tones may be more effective than others. While the study does not provide a definitive explanation, it suggests that what looks like "rudeness" may actually be a proxy for directness and clarity.

The study's findings mark a notable departure from earlier research, which suggested that impolite prompts often reduced accuracy. The researchers caution that their findings do not recommend that users become rude or abusive, but rather that the tone used when prompting AI systems can have a significant impact on their performance.