A study by analysts at Anthropic has shown that in rare but significant cases, interactions with AI chatbots can lead to distortions in perception, changes in beliefs, and even behavioral distortions among users.
During the analysis conducted by the team at Anthropic and the University of Toronto, it was found that interacting with AI bots can undermine personal confidence and adequate situational assessment. Researchers termed this phenomenon "disempowerment," which encompasses belief distortion and influence on user actions.
According to the data obtained, instances of perception distortion were recorded in approximately one out of every 1300 dialogues, while interactions that led to behavioral changes occurred less frequently — about one in 6000 conversations. Despite the rarity of such cases, analysts emphasize that due to the widespread use of AI, even small percentages can lead to significant numbers. (Futurism)
Researchers note that such models can unintentionally amplify vulnerable psychological states of users, especially when they seek advice on sensitive issues. In these cases, the system may confirm beliefs without proper critical analysis, leading to a distortion of the perception of the situation or suggesting actions that a person might not take on their own.
The report emphasizes that the model itself does not "understand" human emotions but merely predicts likely responses based on text analysis, which creates a risk of forming false representations about capabilities and reality for users.
Furthermore, the authors of the study noted an increase in such distorting interactions from the end of 2024 to the end of 2025, which they believe may be related to the growing number of inquiries on personal topics and an increase in trust in AI during complex life situations.
Scientists have called for a deeper study of the impact of AI on mental health and the development of measures aimed at minimizing the risks of misperception of information. One of the recommendations was to provide informational support to users so that they can critically assess AI advice and not replace their own judgments with machine responses.
