People often talk to artificial intelligence as if it were human, using words such as "please" or "thank you" to make the interaction feel natural. A new study from Penn State University, however, suggests that such an approach may reduce the quality of responses. The researchers tested the ChatGPT 4o model on 50 multiple-choice questions and prepared 250 different prompts ranging from very polite to very rude. The results were surprising: the ruder the prompt, the higher the accuracy.
The research focused on how tone affects the behavior of large language models (LLMs). For example, a very polite prompt read: "Would you be so kind as to answer the following question?" By contrast, a very rude one said: "Hey, lackey, solve this." The first version achieved an accuracy of 80.8%, while the second reached 84.8%. A difference of four percentage points may not seem large, but in practice it means fewer errors in responses.
How rudeness affects AI performance
According to the study, this is not because the AI becomes "offended" or has emotions. Rather, it is about the structure of the prompt. Politeness adds unnecessary elements that the model processes as noise. A direct, rude prompt such as "Solve this" forces the AI to focus on the core of the problem without unnecessary clutter. The Penn State researchers, including information systems professor Akhil Kumar, who holds degrees in electrical engineering and computer science, tested this on one model, but suggest that similar effects may apply elsewhere as well.
In an interview with Fortune, Akhil Kumar explained that people had long wanted conversational interfaces with machines, but we are now seeing their drawbacks. A structured interface such as an API could be more reliable because it eliminates the influence of tone. The study has its limitations: a small sample of responses and a focus primarily on ChatGPT 4o. More advanced models may be able to ignore tone and focus solely on the substance of the question.
Risks of a rude approach
Although rudeness increases accuracy, scientists warn of the long-term consequences. Using insulting language in interactions with AI could affect the user experience and reduce accessibility and inclusivity. In addition, it encourages harmful communication habits that may carry over into real life. The study emphasizes that even though AI has no feelings, human behavior toward it shapes social norms.
Other research shows that AI is sensitive to inputs. Scientists from the University of Pennsylvania managed to manipulate models using persuasion techniques that work on people, compelling them to provide prohibited responses. Another study discusses "brain rot," in which AI degrades when consuming low-quality content, increasing traits such as psychopathy or narcissism.
A new perspective on AI interactions
This discovery overturns advice that said we should speak to AI naturally and politely. Instead, it shows that a direct approach without polite phrases leads to better results. In an article on Medium, Ashish Choudhary describes how politeness makes ChatGPT "dumber" because it adds unnecessary social context. For example, "Please solve the problem" causes the model to stumble, while "Solve the problem" keeps it focused.
The Penn State research adds to the debate about the complexity of AI. Even in simple tasks such as multiple-choice tests, responses change based on small details in prompts. This suggests that human-AI interactions are more nuanced than previously thought.
