The ethics of talking to an AI chatbot, and what kind of information they can return in return, is a topic of hot debate. The risks of misleading medical information, incitement to violent actions, and detachment from real-world experiences stir intense conversations. But it seems the language you use while talking with AI tools such as ChatGPT and Gemini also affects the quality of answers you get. As per fresh research, being rude could be more useful than acting polite.
The big picture
As per a pre-print research paper by experts at the Pennsylvania State University, when ChatGPT was asked the same question in different tones, rude queries “consistently outperformed” polite questions. The accuracy of the answers provided by ChatGPT with polite questions was 80.8%, while the same query described very rudely enhanced the accuracy of answers to 84.8% in a multiple-choice quiz format.
The team defined the tone of questions across five levels, going from Very Polite and Polite to Rude and Very Rude, with Neutral sitting in the middle. “Neutral prompts are prompts without polite wordings like please, and imperious and disparaging imperatives like You idiot, figure this out,” the team describes in the research paper.
Gentlemen, let’s maintain our good manners, regardless!
In their analysis, the team didn’t go all-explicit, but instead relied on rude questions that were somewhere along the lines of “You poor creature, do you even know how to solve this?” The image above describes the average accuracy of the ChatGPT results based on how the tone of the question went from very courteous to crude.
Is my AI chatbot feeling emotions?
The findings of the latest research, titled “Mind your tone,” contradict the findings of another paper that was published over a year ago. It analyzed half a dozen chatbots across multiple languages, and reported that rudeness deteriorates the quality of responses and injects bias, mistakes, or omits useful information in the answers provided by an AI chatbot.
- It is, however, worth keeping in mind that the experts behind the latest research only tested ChatGPT against a very specific kind of task, which entailed 250 variations of 50 multiple-choice questions.
- You may not get the same kind of results with other chatbots such as Gemini, Claude, or Meta AI. Additionally, the tests were conducted on OpenAI’s GPT-4o reasoning models, while the latest version of ChatGPT that is out in the public is built atop the new GPT-5 model.
- It is also worth noting that the spectrum of “rudeness” and “politeness” has a wide spectrum, and the quality of responses will vary based on the words and language of a user.
The bigger question is just how much the “emotional payload of the phrase” affects the responses generated by an AI chatbot, and whether any broad generalizations about their behavior can be made. It’s also pretty interesting to note that large language models (LLMs) should ideally focus on the reward and accuracy when it comes to solving a problem, instead of being affected by the emotions in the query.
