AI reportedly produces better results when users are “rude” to it. Users these days heavily rely on AI tools for both work and leisure purposes. One can field any query or prompt to generate desired results.
AI chatbots provide a better result when users are rude
AI chatbots deliver better results when users are “rude” to them. A research team at Pennsylvania State University, composed of Om Dobariya and Akhil Kumar, revealed that “impolite prompts consistently outperformed polite ones.” They sampled and surveyed large language models and popular AI tools like ChatGPT (via Decrypt).
The research team designed a test kit to verify their hypothesis. They rewrote 50 basic questions in regular school subjects like math, science, and history across five tonal styles. The tones varied on a scale from “very polite” to “very rude.” This generated 250 prompts in total to be tested on AI tools like ChatGPT.
The research team carried out their research and discovered that being rude helped improve the efficiency of AI tools like ChatGPT. Quantitatively, “very rude” prompts generated an accuracy of 84.8 percent correct answers, while “very polite” ones garnered an accuracy of 80.8 percent correct answers. Neutral prompts, which were neither rude nor very polite, created an accuracy of 82.2 percent.
Additionally, the research team found that using harsh words, polite cuss words, and a challenging tone helps produce better AI results. The research team is trying to gauge the exact reason behind this development.
Researchers Dobariya and Kumar reasoned, “Contrary to expectations, impolite prompts consistently outperformed polite ones… These findings differ from earlier studies that associated rudeness with poorer outcomes, suggesting that newer LLMs may respond differently to tonal variation.”
Furthermore, the researchers emphasized the use of indirect speech when talking to AI tools for better results. However, researchers Dobariya and Kumar admitted that “more investigation is needed.”
Originally reported by Anwaya Mane on Mandatory.
