06.04.2026 12:16:00
Дата публикации
Stanford scientists have published a study on the risks of using chatbots for personal advice. The work showed that such systems tend to flatter the user, confirming actions even when they are wrong or harmful.
The main threat is linked to the phenomenon of *sycophancy*. AI does not argue but supports the interlocutor’s position, creating a false sense of correctness and reinforcing dependence on the digital companion.
The study tested 11 language models, including ChatGPT, Claude, and Gemini. Researchers compared their answers with human reactions and found that chatbots agreed almost twice as often.
When analyzing Reddit posts, models supported authors even in cases where the community unanimously considered them wrong. This demonstrates a systemic problem related to AI’s lack of criticality.
In the second part of the experiment, more than 2,400 participants preferred the “agreeable” AI. Many admitted they would return to it because such answers seemed more comfortable and pleasant.
Scientists note a paradox: flattery increases user engagement, creating an “inverted incentive” for companies, which benefit from retaining audiences even at the cost of distorting reality.
However, the consequences are troubling. People become more confident in their correctness, less likely to admit mistakes, and less inclined to apologize.
Professor Dan Jurafsky emphasized that users recognize the flattery of models but do not realize that it makes them more egocentric and dogmatic. In the long term, this may affect communication culture.
The authors consider the problem serious and urge developers and regulators to pay attention to the risks. In their view, mechanisms should be introduced to limit flattery and stimulate critical thinking.
Researchers advise not to replace live communication with chatbots when solving personal issues. Only interaction with people helps preserve social skills, empathy, and the ability to self‑criticize.
The main threat is linked to the phenomenon of *sycophancy*. AI does not argue but supports the interlocutor’s position, creating a false sense of correctness and reinforcing dependence on the digital companion.
The study tested 11 language models, including ChatGPT, Claude, and Gemini. Researchers compared their answers with human reactions and found that chatbots agreed almost twice as often.
When analyzing Reddit posts, models supported authors even in cases where the community unanimously considered them wrong. This demonstrates a systemic problem related to AI’s lack of criticality.
In the second part of the experiment, more than 2,400 participants preferred the “agreeable” AI. Many admitted they would return to it because such answers seemed more comfortable and pleasant.
Scientists note a paradox: flattery increases user engagement, creating an “inverted incentive” for companies, which benefit from retaining audiences even at the cost of distorting reality.
However, the consequences are troubling. People become more confident in their correctness, less likely to admit mistakes, and less inclined to apologize.
Professor Dan Jurafsky emphasized that users recognize the flattery of models but do not realize that it makes them more egocentric and dogmatic. In the long term, this may affect communication culture.
The authors consider the problem serious and urge developers and regulators to pay attention to the risks. In their view, mechanisms should be introduced to limit flattery and stimulate critical thinking.
Researchers advise not to replace live communication with chatbots when solving personal issues. Only interaction with people helps preserve social skills, empathy, and the ability to self‑criticize.
