Politeness Makes AI Respond Differently — An Interesting Experiment

Researchers from the University of Innsbruck simulated about 18,000 dialogues with three language models and found that a polite communication style leads to more detailed and informative AI responses, while rudeness reduces the amount of useful information.

Researchers from the University of Innsbruck conducted an experiment, simulating about 18,000 dialogues with three language models. Queries were composed in five styles — from polite and elaborate to short and rude. The results showed that polite communication, reminiscent of ordinary human conversation, prompts AI to generate more detailed responses and provide more useful information. With a polite style, responses could be nearly twice as long, and the amount of useful information increased by about one and a half times. With rude communication, models spent more time processing the query, but the final response was shorter and contained less information, and such results more often required additional verification. The authors attribute the effect to the peculiarities of training models on texts created by humans. The study also revealed that longer responses require more computational resources.