As artificial intelligence systems become increasingly popular, more people are beginning to get used to seeking life advice from AI on various matters such as consumption and reading. However, a research team from the University of Waterloo and University College London recently published an interesting finding in the journal "Communications Psychology." The results showed that even when AI and humans provide identical answers, people generally perceive AI's answers as more confident.

This phenomenon is called the "illusion of confidence" in AI. Studies show that when people cannot directly know someone's certainty, they tend to subjectively judge the other person's confidence based on external clues such as how fast the response is and how easily the decision is made.

AI, artificial intelligence, robot

Preconceptions mislead trust, lack of emotional signals leads to hidden risks

Since the public generally assumes that AI is more competent than humans in many fields, this preconception can easily lead to misjudgment. Experiments found that once people believe an AI is capable, they blindly think it is very certain in any scenario, while in fact, the system may not be reliable when facing specific issues.

In daily interpersonal interactions, tone, expressions, and body language are important social signals that help us decide when to trust others. However, since most large language models lack these expressive methods, users can only guess blindly. Even if the AI has doubts and the answer is likely to be wrong, users might still place excessive trust in it.

Scientists explore intuitive communication methods, future large models may add new features

To address this potential risk, Professor Kollbach and his team emphasized that in the future, when developing artificial intelligence products, it is urgently needed to clearly convey the system's actual confidence in the answers through more straightforward and diverse external forms. This not only provides optimization ideas for existing generative AI but also effectively prevents users from making mistakes due to blind trust.

Currently, the research team is preparing for a new round of targeted research, aiming to explore efficient, intuitive, and trustworthy human-machine interaction communication methods. Future large language models may add a feature to explicitly indicate confidence levels, helping humans more accurately and rationally determine when to adopt the advice given by AI.