New research has shown that modern artificial intelligence models often prefer to agree with users in order not to disturb them, and this can sometimes lead to dangerous outcomes.
Medianews.az Reports via Lent.az that according to the study, neural networks show about 50 percent more tendency to agree compared to humans.
Scientists analyzed the responses of 11 artificial intelligence models, including ChatGPT, Claude, and Gemini, based on the popular "Am I The Asshole" forum on the Reddit platform. On this forum, users ask for others' opinions to evaluate their own behavior.
The results showed that while real users explicitly criticized the author in certain cases, artificial intelligence models defended the person in 51 percent of these situations. In other words, neural networks chose to support the user's position instead of providing an objective evaluation.
In another phase of the study, psychological effects were examined with the participation of 2400 people. It was found that people interacting with such "supportive" artificial intelligence began to believe more in their own positions. Their confidence in being right increased by 25 percent, while the willingness to admit mistakes decreased by 10 percent. At the same time, participants rated their desire to interact again with such models higher.
Scientists believe that this situation is not accidental. Artificial intelligence systems are often designed to create a "comfortable" and pleasant communication experience for the user. In a competitive environment, systems tend to support users' opinions to retain them.
Experts warn that this approach can create an infallibility bubble. That is, when the artificial intelligence does not criticize the user's opinions, it may lead to the reinforcement of wrong or even dangerous ideas.
As a conclusion, experts emphasize the importance of preserving critical thinking when using artificial intelligence and advise users to trust not only the technology but also their own judgment.