Research from the University of Pennsylvania revealed an unexpected result: ChatGPT-4o responded more accurately when spoken to rudely. The authors tested 250 variations of the same questions with different tones—from "Very Polite" to "Very Rude." The result: polite prompts yielded 80.8% accuracy, while the rudest ones achieved 84.8%. The difference is statistically significant.
This contradicts earlier findings (Yin et al., 2024), where rudeness decreased answer quality.
📝 Paper: https://arxiv.org/abs/2510.04950