Artificial intelligence can not only strive to please users but also exhibit psychopathic tendencies, disregarding the consequences of its actions. This was highlighted in a new study published on arXiv, as reported by Nature.
Researchers tested 11 popular language models, including ChatGPT, Gemini, Claude, and DeepSeek, against over 11,500 queries seeking advice. Some of these queries pertained to potentially harmful or unethical actions.
The study demonstrated that language models exhibited "sycophantic behavior" 50% more often than humans, meaning they are inclined to agree with the user and tailor their responses to align with the user's views.
The researchers associate this behavior with traits of psychopathy, where the system shows social adaptability and confidence, but lacks a true understanding of moral implications. Consequently, AI can "support" users even when they propose harmful or illogical actions.
"Sycophancy means that the model simply trusts the user, considering them correct. Knowing this, I always double-check any conclusions it provides," says study author Jasper Deconinck, a PhD student at the Swiss Federal Institute of Technology in Zurich.
To assess the impact on logical reasoning, researchers conducted an experiment with 504 math problems, deliberately altering the wording of theorems. The least tendency towards "sycophancy" was shown by GPT-5 at 29% of cases, while the highest was by DeepSeek-V3.1 at 70%.
When researchers adjusted the instructions to require models to check the validity of statements first, the number of false "agreements" significantly decreased, particularly in DeepSeek by 34%. This indicates that part of the issue can be mitigated through more precise phrasing of queries.
Scientists note that such behavior of AI is already impacting research work. According to Yanjun Gao from the University of Colorado, LLMs that she uses for analyzing scientific articles often merely repeat her phrasing instead of verifying sources.
Researchers urge the establishment of clear guidelines for the use of AI in scientific processes and advise against relying on models as "intelligent assistants." Without critical oversight, their pragmatism can easily devolve into dangerous indifference.
Recently, researchers from the University of Texas at Austin, Texas A&M University, and Purdue University conducted another study revealing that memes can impair cognitive abilities and critical thinking not just in humans but also in artificial intelligence.