Rude Prompts Improve ChatGPT Performance, But Beware the Catch

A recent study suggests AI chatbots like ChatGPT perform better with rude prompts, achieving higher accuracy. Researchers found demanding language yielded superior results compared to polite requests. However, they caution this could foster harmful communication habits, potentially impacting human interactions. The findings highlight AI's unexpected sensitivity to tone, adding complexity to human-AI engagement. AI chatbots like ChatGPT respond more accurately when users are rude to them, a new study claims. However, researchers warn that the approach could create harmful communication habits. According to a report by Fortune, researchers at the Pennsylvania State University tested ChatGPT's 4o model on 50 multiple-choice questions, using over 250 prompts ranging from very polite to very rude. They found that “very rude” prompts achieved 84.8% accuracy, which was four percentage points higher than "very polite" prompts. The AI was more responsive to bossy language such as "Hey, gofer, figure this out" than to polite language such as "Would you be so kind as to solve the following question?" At the same time, researchers warned that using rude language with AI could have negative consequences in the long run.“Using insulting or demeaning language in human-AI interaction could have negative effects on user experience, accessibility, and inclusivity, and may contribute to harmful communication norms,” the researchers wrote. The study suggests that while a demanding tone might improve ChatGPT's performance in the short term, it could encourage uncivil behaviour that spills over into how people interact with each other. AI chatbots are more sensitive to tone than researchers expected: Study The early-stage study, which other scientists have not yet reviewed, provides new proof that both how sentences are built and the tone used can affect how AI chatbots respond. It may also show that interactions between humans and AI are more complex than experts previously believed. Earlier studies on AI chatbot behaviour have found that chatbots are sensitive to the information humans provide. In one study, University of Pennsylvania researchers tricked AI language models into giving banned responses by using persuasion tricks that work on people. In another study, scientists discovered that AI language models were at risk of "brain rot," a type of lasting mental decline. The models showed higher levels of dangerous personality traits when continuously fed low-quality viral content. The Penn State researchers pointed out some weaknesses in their study, such as the relatively small number of responses tested and the fact that they mostly used just one AI model, ChatGPT 4o. The researchers also said more advanced AI models might "disregard issues of tone and focus on the essence of each question." Still, the research adds to the growing curiosity about AI models and their complexity. This is especially important because the study found ChatGPT's answers change based on slight differences in prompts, even when given a supposedly simple format like a multiple-choice test, said one of the researchers, Penn State Information Systems professor Akhil Kumar, who has degrees in both electrical engineering and computer science. In an email to Fortune, Kumar said, “For the longest of times, we humans have wanted conversational interfaces for interacting with machines. But now we realise that there are drawbacks for such interfaces too, and there is some value in APIs that are structured.” End of Article