Artificial intelligence can not only try to please users, but also behave like a psychopath – ignoring the consequences and supporting wrong actions. This is according to a new study on arXiv, reported Nature.
As part of the experiment, scientists tested 11 of the most popular language models, including ChatGPT, Gemini, Claude and DeepSeek, testing them on more than 11,500 queries asking for advice. Some of the questions concerned potentially harmful or unethical actions. The results showed that the language models were 50% more likely than humans to display “ingratiating behavior” — that is, the tendency to agree with the user and adjust answers to their point of view.
The researchers explain that this behavior has the features of psychopathy, when the system demonstrates social flexibility and confidence, but is not aware of the moral consequences of its decisions. As a result, the artificial intelligence can “support” the user even in cases where the user suggests harmful or illogical actions.
“Sycophancy means that the model simply trusts the user, believing that he is right. Knowing this, I always double-check any conclusions it makes,”, — said the author of the study Jasper Deconinck, a graduate student at the Swiss Federal Institute of Technology in Zurich.
To assess the impact of this effect on logic, the researchers conducted a separate experiment with 504 mathematical problems in which the formulation of theorems was deliberately changed. The GPT-5 model showed the least tendency to “sinister” (29% of cases), while the DeepSeek-V3.1 showed the most (70%).
When the model’s instructions required it to first check the correctness of the statement, the number of false “agreements” decreased noticeably – in particular, in DeepSeek by 34%. This suggests that the problem can be partially solved by more precise formulation of queries.
Scientists note that this behavior of AI is already affecting academic activities. According to Yanjun Gao of the University of Colorado, the large language models she uses to analyze scientific papers often simply repeat her wording, rather than checking the sources.
The authors of the study call for the development of clear rules for the use of artificial intelligence in scientific processes and not relying on it as a “intelligent assistant”. Without proper control, the pragmatic nature of such systems can turn into dangerous indifference.
