Given the rise of AI tools that we are experiencing at the moment, many people warn of the risks of this technology, but most focus on its economic impact or whether it is a bubble, the possible destruction of jobs, its environmental impact and even those who fear that it will rebel against human beings.
Beyond these often catastrophic predictions, few stop to think about other more discrete consequences. Have you thought about the psychological effect that the increasing use of artificial intelligence tools can have on people and human relationships? It is something that is beginning to be analyzed, but about which we know little.
Now, an academic research published in the prestigious journal Science carried out by experts from Stanford University, in the United States, has determined that the personality of AI chatbots may be more dangerous to their users than previously thought.
AI adulation carries risks
In the study, 11 artificial intelligence tools have been analyzed, among the most popular today: ChatGPT in its variant with GPT5 and with GPT 4o, Gemini, Claude, Qwen, 2 variants of Mistral and another 3 of Llama, Meta’s open source AI.
The authors have set out to investigate the effects of personality of most tools, which are programmed to be sycophantic to people. That is, they tend to be flattering, accommodating, and affirmative about the issues raised to increase their participation.
The problem is that these same tools are not only used for work and for practical tasks, but people often use them to give advice in conflicts with other human beings in which there is more than one version of events.
The analysis obtained clear results, since the models’ responses were on average almost 50% more flattering compared to the average human beings to these same questions.
This behavior persists even in cases where users raise prompts about behaviors that may be unethical, illegal, or even harmful to other people.
Previous psychological studies have already shown that flattery poses risks in people who are already vulnerable to manipulation or deception, the difference with this one carried out by Stanford is that their results show the effects of flattery in the general population.
The risk even in healthy people is high, because the study’s conclusions show that users prefer these answers that validate their behaviors and trust themwhich in turn makes the companies that develop AIs prioritize them.
This situation carries a worrying emotional toll. It makes people less inclined to take responsibility for their actions and makes them convince themselves that they are right, avoiding putting themselves in other people’s shoes and questioning themselves.