Core Findings - A surprising discovery from researchers at Penn State University indicates that more polite questions lead to less accurate responses from ChatGPT, with an average accuracy of 80.8% for very polite inquiries compared to 84.8% for very rude ones [3][4]. Summary by Sections Experiment Design - The research team focused on ChatGPT-4o, constructing a dataset of 50 multiple-choice questions with five variations of politeness [5]. - Each question was rewritten in five tones: very polite, polite, neutral, rude, and very rude, covering subjects like math, science, and history [6][7]. Results - The results showed a consistent increase in accuracy from very polite to very rude tones, with statistical significance (p≤0.05) across eight comparisons [8]. - The phenomenon termed "counterintuitive tone effect" suggests that ChatGPT-4o performs better with direct commands than with polite requests [8][11]. Implications - The findings challenge the conventional understanding of human interaction, where politeness is associated with cooperation, indicating that in machine interactions, directness may yield better results [9][11]. - The research suggests that the model's response to politeness is not emotional but algorithmic, with polite phrases potentially adding unnecessary complexity that hinders performance [9][10]. Future Directions - Initial tests on other models like Claude and GPT-3 indicate a trend towards reduced sensitivity to tone, suggesting future architectures may focus more on content rather than expression [12].