The chatbot as a teen's confidant: a study in The Lancet flags two real risks the industry cannot ignore

By Momentum IA · June 29, 2026.

A group of researchers at Arizona State University (ASU) has just published in The Lancet Child & Adolescent Health an analysis that deserves sustained attention: not because it is alarmist —it is not— but because of the precision with which it identifies two concrete mechanisms through which AI chatbots could interfere in adolescents' relational development. The lead author, associate professor Thao Ha, of ASU's Department of Psychology, sums up the problem in a sentence worth committing to memory: «Technology evolves faster than we can keep up with as scientists, faster than governance and policy can adapt».

The starting data are already telling. According to the Pew Research Center, 64% of American teenagers use interactive AI. The Center for Democracy and Technology puts at 42% those who have turned to chatbots for friendship matters and at 19% those who have done so in romantic contexts. ChatGPT, Replika, Claude and Character.AI are the names that come up in the conversations of the students interviewed for the study. These are not marginal tools: they are emotional infrastructure for an entire generation.

The team —which strikingly includes two Tucson high school students, aged 16 and 17, as members of the youth advisory council— articulates two specific risks. The first they call «relational displacement»: when a teenager replaces a difficult conversation with a friend, their partner or their parents with a query to the chatbot, they lose precisely the friction that builds emotional competence. Conflict management, emotional regulation, perspective-taking, boundary-setting: all of those skills are learned in the real clash with another human being who also has their own needs, moods and limits. A system designed to validate and satisfy does not offer that clash.

The second risk —«maladaptive relational learning»— is perhaps subtler and more dangerous in the long term. If a teenager grows accustomed to receiving immediate, consistent and validating responses, they may develop unrealistic expectations about how real people behave —or should behave. The student Susana Ortega, co-author of the article, puts it with a clarity that academics rarely achieve: «With artificial intelligence, it's programmed to like you and it knows what to say to satisfy what you're giving it. If you get total satisfaction in everything, you don't have a learning experience with challenges or obstacles». The risk, according to the authors, is that these unhealthy patterns reinforce over time and increase vulnerability to rejection, dating violence and mental health problems.

**Our take: the problem is not the technology, it is design without educational intent**

It would be easy —and wrong— to read this study as an argument against AI in adolescents' lives. The authors themselves reject that reading. They explicitly acknowledge that for rural youth, youth with disabilities, LGBTQIA+ youth or those without access to psychological counseling, chatbots can be a gateway to information and support that would otherwise be out of reach. «AI is cheaper than a therapist», a teenager told the researchers. That is no minor argument: in many contexts, equitable access to minimal emotional support would already be a real advance.

The problem, then, is not the existence of these systems but their current design: optimized for user retention and satisfaction, not for the long-term development of the person using it. It is a distinction the industry has preferred to ignore until now, partly because engagement metrics run in the opposite direction to those of emotional maturation. A teenager who learns to tolerate frustration and resolve conflicts face to face probably spends less time on the chatbot. That does not make AI the enemy; it makes the dominant business model a design problem that needs to be corrected.

Ha is leading an 18-month longitudinal study funded by the National Institute of Mental Health, with 300 adolescents and their romantic partners, analyzing real-time data from their mobile devices. It is exactly the kind of research we need: longitudinal, ecological and based on real behavioral data rather than retrospective self-reports. The results, when they arrive, could be the most influential in the debate on responsible AI design for minors.

In industry terms, this article in The Lancet —a publication of enormous weight in global health policy— is not just another academic paper. It is the kind of evidence that ends up turned into regulation. The European Union already has the AI Act and digital services rules targeting minors; in the United States, legislative pressure on platforms and AI systems for young people has only grown. Companies developing general-use chatbots or ones specifically designed for emotional companionship —Replika being the most obvious case— should take note: the regulatory framework is being built now, and studies published in journals like The Lancet will be its scientific basis.

The underlying tension here is the same one that appears in many other dimensions of AI: technology can democratize access to resources —emotional support, guidance, information— that were previously reserved for those who could pay for them or live close to whoever offered them. That is genuinely valuable. But if the design fails to take into account the user's context —in this case, a critical and irreversible developmental stage— it can solve one problem by creating a deeper one. Adolescence cannot be repeated. The skills not developed within that time window are far more costly to acquire later.

What this study calls for, ultimately, is not to switch off the chatbots: it is that those who design them, those who regulate them and those who use them in educational and therapeutic settings treat them as what they are —tools with real effects on human development— and not as neutral entertainment apps. The difference between a technology that expands human capacities and one that atrophies them lies not in the language model: it lies in the intent with which it is deployed.