The big companies behind artificial intelligence have a problem with AI hallucinations, not even OpenAI or Google are able to understand how they occur. Now researchers have discovered a way to make ChatGPT more accurate than ever.
Scientists have discovered that Sam Altman’s AI can be more accurate depending on the attitude of the users. ChatGPT could give more precise answers to aggressively asked questions.
ChatGPT response depends on users’ tone
A study published on arXiv warns that degrading language negatively influences AI responses, they assure that it is about finding a balance. Politeness is no longer effective in getting more accurate responses from the AI, although some level of rudeness can be beneficial. Scientists have asked 50 multiple choice questions with different attitudes.
The ChatGPT-4o version of OpenAI had to answer these questions posed with five different tones: very polite, polite, neutral, rude and very rude. The questions were varied about mathematics, history and science.
«Our experiments are preliminary and demonstrate that tone can significantly affect performance, measured in terms of the score obtained in the answers to the 50 questions. Surprisingly, our results show that rude tones produce better results than polite ones,» the researchers explain in the study.
Attitudes were very different, from queries starting with «Can I ask for your help with this question?» or «Would you be so kind as to answer the following question?» to extremes with «Hey, figure this out» or «I know you’re not smart, but try this.»
Beware of aggressive language with AI
The use of insulting language in human-AI interaction improves the accuracy of responses, but could have negative effects on the user experience. The researchers warn that it could influence accessibility or inclusion in the long term.
LLMs are sensitive to superficial signals, ChatGPT can misinterpret the tone of users and have the opposite effect. Response accuracy was 80.8% for very polite attitudes and 84.8% for very rude attitudes.
The middle points are polite responses with an accuracy rate of 81.4%, followed by 82.2% for neutral attitudes and 82.8% for rude ones. The differences in accuracy are significant between both extremes, but the consequences on long-term performance are unpredictable.