Rudeness Towards AI Increases Response Accuracy by Four Percent

Politeness may be a virtue in human relationships, but when it comes to communication with artificial intelligence, effectiveness could be quite the opposite. The latest research from Penn State University presents unexpected results that could change the way we interact with AI assistants.

The study ‘Mind Your Tone’ by Oma Dobariye and Akhil Kumar shows that impolite queries consistently yield better results than polite ones when asking large language models. Specifically, ‘very rude’ queries produced accurate responses in 84.8 percent of cases, while ‘very polite’ ones achieved 80.8 percent accuracy.

A Turnaround from Previous Findings

These results directly contradict earlier research from 2024 that suggested AI models mimic human social norms and reward politeness. The study ‘Should We Respect LLMs?’ concluded that impolite queries often reduce model effectiveness, while excessive politeness did not provide clear advantages.

Penn State researchers conducted an extensive experiment with 250 prompts generated by rephrasing 50 basic questions from the fields of mathematics, science, and history. Each question was formulated in five tonal levels, from ‘very polite’ to ‘very rude’, and then posed to the ChatGPT-4 model.

Why Directness Wins?

One possible explanation lies in the very structure of polite messages. Phrases like ‘Could you please tell me…’ introduce additional linguistic complexity and potential ambiguity. In contrast, the direct request ‘Tell me the answer’ eliminates superfluous elements and provides AI with a clearer intent.

Thus, the same words that facilitate human communication can complicate machine logic. Newer AI models seem to no longer function as social mirrors but as strictly functional systems that prioritize directness over politeness.

A Paradigm Shift

The research supports recent findings from the Wharton School of Business regarding the importance of prompt engineering, or the skill of precisely formulating queries to achieve better AI results. Tone, long considered insignificant, is (at least for now) proving to be almost as important as the choice of words themselves.

Meanwhile, California has become the first U.S. state to introduce explicit rules for AI ‘companion’ chatbots, requiring them to identify as artificial intelligence and limiting discussions on sensitive topics with minors.

Although the study has not yet undergone peer review, it is already sparking discussions among experts who see it as a sign that future models may need social calibration, not just technical refinement. OpenAI CEO Sam Altman previously warned that saying ‘please’ and ‘thank you’ to ChatGPT is a waste of time and money, and it seems he was right.