Researchers from Pennsylvania State University conducted an intriguing study to examine whether the tone of questions affects the responses generated by AI, specifically GPT-4o. They posed 50 inquiries across various domains, varying the phrasing from extremely polite to overtly rude. The results revealed a surprising trend: using a more aggressive tone increased the accuracy of responses from 80.8% to 84.8%. However, this study faced criticism for its limited sample size. In response, other researchers carried out a broader investigation, testing multiple AI models with hundreds of questions each. Their findings indicated that Gemini 2.0 Flash was largely unaffected by the tone of the inquiries. Conversely, both GPT-4o mini and Llama 4 Scout showed a decline in performance on humanities-related tasks when faced with rude prompts. Efforts to establish a consistent prompting style for AI have been ongoing, yet the results remain unpredictable. Previous studies suggested that models like GPT-3.5 and Llama2-70B tend to respond more favorably to polite prompts, with accuracy significantly dropping when confronted with rudeness. Some researchers from the University of Pennsylvania took this further, noting the difficulty in predicting how tone influences responses, as sometimes politeness can enhance results while at other times it can hinder them. The takeaway is clear: there is currently no definitive guideline indicating whether being rude or polite yields better AI performance. With new models emerging at a rapid pace, researchers struggle to keep up with systematic evaluations of their behavior. For now, the best approach is to experiment and determine what works best in your interactions with AI. How does your communication style affect AI responses?
Informational material. 18+.
