A new study from Penn State has revealed that ChatGPT's accuracy may improve when it's prompted rudely, according to Fortune Magazine. Researchers tested OpenAI's GPT-4o model using over 250 unique prompts ranging from very polite to very rude and the results were unexpected.
According to the preprint study, which has not yet been peer-reviewed, the AI model performed best when given blunt or impolite instructions. In fact, "very rude" prompts like "Hey, gofer, figure this out," led to an 84.8% accuracy rate on multiple-choice questions-four percentage points higher than polite phrasing like "Would you be so kind as to solve the following question?"
The researchers believe this could indicate that AI models respond differently depending on tone and prompt structure, suggesting that AI-human interaction is more complex than previously thought. "Minor changes in how a question is asked can significantly affect the outcome," said study co-author and Penn State professor Akhil Kumar.
However, the study also warned of potential risks. Rude prompts may encourage harmful communication habits and affect accessibility and inclusivity in AI use. Despite the performance boost, researchers caution against normalizing "uncivil discourse" during AI interactions.
Previous studies have shown that AI chatbots are sensitive to the quality and tone of inputs. In some cases, constant exposure to low-quality content has even led to what researchers call "brain rot"-a degradation in output quality over time.
"For the longest of times, we humans have wanted conversational interfaces for interacting with machines," Kumar told Fortune in an email. "But now we realize that there are drawbacks for such interfaces too and there is some value in APIs that are structured."
The Penn State study highlights the need to understand not just what we ask AI, but how we ask it-raising ethical questions about the future of human-AI interaction.
