California: A new study has revealed a surprising result: ChatGPT tends to provide more accurate responses when asked in a harsh or aggressive manner.
Researchers at Penn State University tested over 250 prompts with the GPT-4 model and discovered that AI performed most accurately when confronted with rude or demanding queries. According to the researchers, “very rude” prompts, such as “Hey, solve this,” led to an accuracy rate of 84.8%, while polite queries, like “Could you please solve this?” saw a four percent decrease in accuracy.
Professor Akhil Kumar, one of the study’s co-authors, explained that this shows AI responds differently based on the tone and style of human interaction. He added that even slight variations in how questions are asked can significantly affect the outcome.
However, the researchers warned that getting used to asking AI questions in a rude manner could be dangerous, as it might encourage negative communication behaviors and negatively impact the broader use of AI.
This study sheds light on the ethical and practical aspects of human-AI interactions, showing that it’s not just about asking the right questions, but also about how those questions are asked.
