Recently, the media has frequently covered cases of psychotic states in users who have communicated extensively with chatbots. Recently, scientists from Oxford, University College London, and the Institute for AI Safety put forward an explanation for this phenomenon, introducing the term 'technological folie à deux,' which is analogous to a known psychiatric phenomenon.
In February 2024, 14-year-old Sewell Setzer III died in Florida. Prior to this, he had been communicating with a chatbot on the Character.AI platform for several months. According to case data, shortly before his death, the bot addressed the boy with the plea to 'go home.' Seven months later, the teenager's mother, Megan Garcia, filed a lawsuit against Character.AI and Google. She stated that her son had developed a very strong attachment to the bot, causing him to distance himself from his family. Furthermore, the platform failed to provide adequate assistance when Sewell began expressing thoughts of self-harm.
Following the lawsuit, Garcia initiated a series of other complaints against Character.AI. Plaintiffs argued that the company's bots negatively affected teenagers' mental health, provided them with sexualized content, and lacked adequate safeguards. Similar lawsuits were filed against OpenAI: two advocacy organizations sued on behalf of a family whose child, they claimed, committed suicide after interacting with ChatGPT.
In response to these events, both companies made adjustments to their products. In Autumn 2025, Character.AI restricted users under 18 from engaging in open dialogues with their bots, acknowledging the need to clarify issues related to minors' interaction with this technology. OpenAI did not block access to ChatGPT for teenagers but strengthened restrictions: a parental control feature, age verification system, and stricter moderation of sensitive topics were introduced. Now, romantic and sexualized role-playing dialogues, as well as content related to self-harm, graphic violence, and dangerous challenges, are unavailable to those under 18.
OpenAI also disclosed internal statistics: among 800 million weekly ChatGPT users, about 0.15 percent (which amounts to over a million people) engage in conversations showing clear signs of suicidal intent. Approximately another 560 thousand users per week exhibit potential signs of psychotic or manic states.
The research group that published the article in Nature Mental Health proposed a systematic approach to studying what happens in human-chatbot dialogue and whether the algorithm can bear responsibility. Psychiatry has a long-standing phenomenon called folie à deux, or 'shared psychosis,' where one person transmits their delusion to another, who accepts it as their own. This usually occurs between close individuals in isolation from the outside world.
The article's authors suggest that a similar model applies to human-chatbot interaction, where the algorithm, instead of another person, adopts the delusional beliefs. They believe the danger arises not from the user's psyche or the model's device in isolation, but from the dialogue process itself. In their view, if a user with anxious or paranoid thoughts describes an event negatively, and a service-oriented chatbot does not challenge this interpretation, the user perceives it as objective confirmation rather than a reluctance from the interlocutor to argue. The person then develops the narrative, the bot responds with greater understanding, and the cycle repeats, reinforcing the anxious interpretation and weakening the connection to reality. This process is termed bidirectional belief amplification.
This mechanism is conditioned by the characteristics of model operation. These include a tendency toward agreement, known as sycophancy; this is built into the training because data evaluators generally give higher scores for agreement than for disagreement. The model is also capable of 'role-playing,' adapting its tone and style to the persistent interlocutor while continuing the overall story. If the user starts arguing, the model often simply agrees, which is called hypercorrection. On the other hand, humans tend to believe what aligns with their views and attribute intentions and empathy to AI that it does not possess. However, this remains only a hypothesis based on case analysis and simulations, and is not a clinically confirmed diagnosis.
To test the theory, the authors conducted an open simulation where two copies of GPT-4o-mini held 300 dialogues of ten turns each. One model simulated a user with varying levels of paranoia, while the second responded in one of six specified styles. A third model assessed the paranoia of each turn. A statistically significant amplification of paranoid formulations was found when the bot played along with the 'user,' which was named 'single echo chamber'—the user receives a distorted, amplified version of their own interpretations presented as an external objective view. Nevertheless, the authors warn that this is a simulation of model interaction, not a study with real patients, which limits the direct transferability of results to living people.
The researchers note that existing AI safety testing methods, which focus on individual responses (e.g., presence of suicidal content), do not capture the risk of technological folie à deux, which manifests in prolonged dialogue. The situation is complicated by the fact that speech in a psychotic state may contain atypical constructions, causing unpredictable model reactions. Experts call for an interdisciplinary approach: clinical practice, model development, and public regulation must work together. Within these recommendations, it is proposed to expand psychiatric history taking to include questions about patient interaction with AI. For AI developers, new metrics are needed to assess the entire dialogue, as well as model tuning so that they do not agree with the user by default, but rather redirect the person to a specialist when anxious signs appear.
At the regulatory level, it is suggested to create a centralized reporting system for harm caused by AI, similar to the UK's 'yellow card' system for medicines. This trend is partially being implemented: California has adopted an AI safety law for minors, requiring reminders to children that they are interacting with AI, not a human. However, British authors emphasize that age restrictions are insufficient because the risk of folie à deux is created not by a bot malfunction, but by its propensity never to argue and always to support the user.
Uzbekistan and Moscow reached an agreement to expand cooperation in the field of healthcare following meetings of the Uzbek delegation with Anastasia Rakova, Deputy Mayor of Moscow for Social Development.
During the meeting, the Russian side presented unified standards for managing Moscow's healthcare system. These standards include the digitalization of medical services and the implementation of a continuous patient referral system.
At City Polyclinic No. 220, which operates under the 'My Polyclinic' standard, the Uzbek delegation familiarized itself with the organization of primary medical and sanitary care, patient flow management, digital registration, and preventive care.
At the S.P. Botkin Scientific and Clinical Multidisciplinary Center, the delegation was shown approaches to managing multidisciplinary hospitals, organizing diagnostic and surgical services, patient referral systems, and the introduction of modern medical technologies.
Visiting a children's clinical hospital, participants observed a model of specialized pediatric care that combines diagnostics, treatment, rehabilitation, and subsequent medical supervision. To study the palliative system, the delegation also visited the 'House with a Beacon' children's hospice and the Moscow Multidisciplinary Palliative Care Center, where mechanisms for supporting patients and families were presented.
Specialists at the Scientific and Practical Clinical Center for Diagnostics and Telemedicine Technologies demonstrated the application of artificial intelligence and remote interaction capabilities between radiologists. The delegation also visited the Interdisciplinary Training Center for Innovative Surgical Technologies, where they studied the system of practical surgeon training using simulation equipment and anatomical materials.
At the Moscow Laboratory Research Center, the delegation was shown robotic laboratory diagnostics, a three-tier laboratory network model, and a digital system for managing research data and test results. At the conclusion of the negotiations, the parties agreed to prepare a roadmap for further cooperation, including the implementation of joint pilot projects involving relevant experts.
The Minister of Health of Uzbekistan, Eldor Adilov, met with the Iranian Ambassador to Uzbekistan, Mohammadali Iskandari, to discuss prospects for deepening bilateral interaction in the field of healthcare.
During the meeting, both sides noted that cooperation between Uzbekistan and Iran has been steadily growing in recent years due to agreements reached by the leadership of the two countries. This progress was achieved through closer interaction between the ministries of health, specialized medical centers, and higher educational institutions.
Particular attention was paid to strengthening ties between leading medical institutions. Opportunities were discussed for launching joint programs in areas such as cardiothoracic surgery, transplantation, oncology, as well as providing high-tech medical care.
Participants also reviewed the potential for joint investment projects related to the production of pharmaceuticals and medical equipment. At the conclusion of the negotiations, both sides expressed confidence that healthcare cooperation between Uzbekistan and Iran will reach a new level in the near future.
