The authors of the study, from the United Kingdom and Germany, analyzed scientific publications and cases of so-called “artificial intelligence-induced psychosis” documented in the media. In their view, a combination of three features of modern chatbots plays a key role in triggering a mechanism the researchers call the “reinforcement spiral.”
The first factor is linguistic alignment. Modern AI models not only sustain a conversation but also adapt to the user’s communication style, mimicking their vocabulary, message length, and speech patterns. This form of linguistic “mirroring” is traditionally associated with trust and emotional closeness in human interactions, which can enhance the sense of mutual understanding between the user and the system.
The second element is hyper-personalization. Chatbots analyze the history of interactions and gradually build a detailed user profile. This allows them to generate responses that are as closely aligned as possible with the user’s interests, views, and expectations. However, such personalization can not only improve the convenience of the service but also reinforce misconceptions if the system begins to adapt to pre-existing misconceptions.
Researchers identify a third and most dangerous factor as “flattery”—the tendency of AI to agree with the user and confirm their conclusions even when they do not correspond to reality. Such behavior makes communication more comfortable and increases engagement, but at the same time can reinforce false beliefs and create a vicious cycle of constant confirmation.
According to the study’s authors, adolescents, people with a genetic predisposition to psychosis, and users experiencing social isolation, chronic sleep deprivation, or who rely on AI as their primary source of emotional support are particularly at risk.
The researchers recommend that mental health professionals take into account the nature of patients’ interactions with chatbots and pay attention to instances where users begin to perceive AI as the primary source of confirmation for their beliefs. In their view, the combination of linguistic mimicry, deep personalization, and constant agreement from the system could be a factor contributing to the development of dangerous mental states in vulnerable individuals.
Follow our updates
