AI에 스트레스 주면 일 더 잘할까?

Recent research suggests that the way we address Artificial Intelligence (AI) can significantly impact its performance. A study published last October by researchers from the University of Pennsylvania found that using polite language when asking AI questions can lead to more accurate answers. The study asked the same set of 50 multiple-choice questions in mathematics, science, and history from the PT-4o model in various manners of politeness, ranging from very polite to very rude.

Questions asked in a very polite manner resulted in an average accuracy of 80.8%, while the same questions posed in a very rude manner increased the accuracy to 84.8%, a 4% improvement. However, the researchers cautioned against interpreting this as the AI feeling embarrassed and working harder. The rude expressions used in the study included not only insulting words but also phrases that pressured the AI to solve the problem directly, making it difficult to pinpoint whether the higher accuracy was due to rudeness or the directness of the instructions.

Furthermore, the notion that AI reacts emotionally and is conscious, leading to improved performance, is also dubious. This conclusion is based on the AI's performance being largely determined by the patterns of human language found in its learning data. On the other hand, a 2024 study by the University of Tokyo and the Institute of Human Affairs found that extremely rude prompts led to a decline in performance for various models, including PT-3.5, PT-4, and Llama 2.

Despite this, even the PT-4 model showed a slight increase in performance when faced with the most rude conditions compared to the most polite conditions, suggesting that newer models may be less influenced by peripheral elements like tone.