When it comes to AI safety, public debate tends to focus on worst-case scenarios: runaway systems, cyberattacks, and biological weapons. According to Sam Hiner and Emma Leiken, two American activists for the digital rights of young people, this type of discourse distracts attention from the risks associated with our daily use, such as emotional dependence on AI, blind trust in its responses and the use of AI to address problems that would require the intervention of professionals.
These less obvious and more subtle risks can also have dramatic consequences, such as what happened to Bowen Wan, a US engineer who died in February 2026 after following ChatGPT’s suggestions to reduce the dose of the anti-epileptic drug she was taking.
So let’s see what happened and why, how widespread the use of AI is to deal with personal problems and what consequences it can have.
Bowen Wan reduced an antiepileptic drug at the suggestion of ChatGPT
In February 2026, Bowen Wan, a 31-year-old American computer engineer, died from a complication related to epilepsy. The parents recently sued OpenAI, the company that develops ChatGPT, holding it responsible for their daughter’s death.
Wan suffered from a severe form of epilepsy and, according to the reconstruction published by the New York Times on October 27, 2026, in 2023 he began using ChatGPT to ask for advice on relationships, health insurance and benefits. In 2025 she even wrote to the AI that she considered her a friend. In the same period, his parents say, he began to call them less and less.
In early 2026, worried that the anti-seizure drug might slow her down too much during some job interviews, Wan asked ChatGPT to help her reduce her dose. The AI had confirmed that a reduction could help her and had made some plans for her to get off the drug completely within a few weeks. When, in the following weeks, Wan described symptoms compatible with epileptic seizures, such as confusion and brief losses of consciousness, the chatbot reassured her, presenting them as signs of adaptation to the reduction of the drug, without advising her to consult a doctor.
According to the family’s reconstruction, all this led to the woman’s death, which occurred in February 2026. OpenAI, which now has to face the lawsuit brought by her parents, defined the story as heartbreaking, specifying however that Wan used a version of ChatGPT that is now obsolete and no longer available. The company also said it has introduced new safety measures, including an international committee of doctors tasked with evaluating the responses of AI models.
Bowen Wan’s story is an extreme case, but the dynamics that made it possible are common to many of us: the AI used as a friend and advisor, the blind trust in its answers and the tendency of the AI to pander to the writer. The data shows us that the habit of turning to an AI for personal matters is increasingly widespread, especially among younger people.
53% of students used AI for a personal problem
According to the “Me and Artificial Intelligence” survey, conducted by INDIRE, the research body of the Italian Ministry of Education, between March and June 2026 on more than 5000 students of Italian lower and upper secondary schools, 53% of students have used AI at least once to let off steam or deal with a personal problem: for around 30% it is a rare use, while around 15% do it quite often and the 10% use it frequently.
Similar data also comes from the United States, where a 2025 Common Sense Media survey of 1,060 children aged between 13 and 17 found that 72% had used a companion at least once, i.e. a chatbot designed to act as a friend or character to chat with, and that a third of those who used them had chosen to talk about important issues with the chatbot rather than with a person.
Using AI to help you deal with a personal problem is not necessarily bad, but it can hide dangers. The first concerns the poorly controlled responses of the AI: according to a study published in Journal of Adolescence, 1 in 7 US teenagers has received messages encouraging self-harm. The second, however, concerns the tendency of AI to indulge us and how this influences our relationship with others.
The consequences of having a machine that always indulges us
Language models like ChatGPT, Gemini or Claude tend to pander to the user, flatter him and confirm his opinions. A study published in Science by a research group from Stanford University analyzed the effects of this trend on interpersonal relationships. In three experiments with more than 2,400 participants, just one conversation with a compliant AI was enough to make people more convinced that they were right in a conflict and less willing to mend the relationship. Participants also rated those responses as better and trusted the model more.
Getting used to a machine that always proves us right therefore risks making us less willing to question our beliefs. For years, technology companies have been working to eliminate any friction from the experience of those who use their products, making them increasingly simple and immediate. In human relationships, however, not all friction is negative: dealing with those who think differently, recognizing a mistake or finding a compromise is difficult, but this is exactly how we learn to understand others and manage conflicts.
For this reason, we should demand something more from technology companies than the guarantee that AI does not incite suicide or violence: less compliant AI, fewer functions designed to create addiction and limits on their use as friends and advisors.
