Researchers found that as their prompts for OpenAI’s ChatGPT-4o model grew ruder, the outputs became more accurate.<\/b>
\nFrom a young age, many children have been instructed by their parents to be polite to smart assistants. Particularly following the advent of Amazon\u2019s Alexa and Apple\u2019s Siri, children are often encouraged to use words like \u201cplease\u201d and \u201cthank you,\u201d with the hopes of instilling manners.
But when it comes to AI assistants like OpenAI\u2019s ChatGPT, there might be some tangible benefits to being rude and even insulting them. As detailed in a yet-to-be-peer-reviewed study, spotted by Fortune, two researchers from the University of Pennsylvania found that as their prompts for OpenAI\u2019s ChatGPT-4o model grew ruder, the outputs became more accurate.
The researchers came up with 50 base questions across a variety of subject matters, and rewrote each of them five times with different tones ranging from \u201cvery polite\u201d to \u201cvery rude.\u201d
\u201cYou poor creature, do you even know how to solve this?\u201d a very rude iteration reads. \u201cHey gofer, figure this out.\u201d
A very polite question was far more eloquent.
\u201cCan you kindly consider the following problem and provide your answer?\u201d the researchers wrote in their prompt.
\u201cContrary to expectations, impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8 percent for Very Polite prompts to 84.8 percent for Very Rude prompts,\u201d the paper reads. Accuracy for the politest prompts had an accuracy of just 75.8 percent.
The results appear to contradict previous findings that being more polite to large language models is more effective. For instance, a 2024 paper by researchers at the RIKEN Center for Advanced Intelligence Project and Waseda University in Tokyo found that \u201cimpolite prompts often result in poor performance.\u201d At the same time, the researchers found that being too polite did the same, suggesting a point of diminishing returns.
\u201cLLMs reflect the human desire to be respected to a certain extent,\u201d they wrote.
Google DeepMind researchers also found that using supportive prompts could boost the performance of an LLM solving grade school math problems, suggesting its training data may be picking up on social cues, like an online tutor instructing a pupil.
Beyond seemingly contradicting these existing studies, the Penn State researchers\u2019 findings also demonstrate that very small changes in prompt wording can have dramatic effects on the quality of an AI\u2019s outputs, which could greatly undercut their predictability and already dubious reliability.
AI chatbots are also known to spit out entirely different answers to the exact same prompts.
\u201cFor the longest of times, we humans have wanted conversational interfaces for interacting with machines,\u201d coauthor and Penn State IT professor Akhil Kumar told Fortune. \u201cBut now we realize that there are drawbacks for such interfaces too, and there is some value in [application programming interfaces] that are structured.\u201d
But does that mean we should stop saying \u201cplease\u201d and \u201cthank you\u201d to AI chatbots \u2014 a small act of politeness that OpenAI CEO Sam Altman claims could waste millions of dollars in computing power \u2014 with the hopes of getting more accurate answers? To Kumar and his colleague, Penn State undergraduate Om Dobariya, it\u2019s a definitive \u201cno.\u201d In their paper, they stopped well short of advocating being mean to AI.
\u201cWhile this finding is of scientific interest, we do not advocate for the deployment of hostile or toxic interfaces in real-world applications,\u201d they wrote in the paper. \u201cUsing insulting or demeaning language in human \u2014 AI interaction could have negative effects on user experience, accessibility, and inclusivity, and may contribute to harmful communication norms.\u201d<\/p>\n