Science Has a Teen AI Problem: Experts Map What Must Be Studied Now

More than two-thirds of American teenagers now use AI chatbots, and roughly one in four uses one every day, yet the scientific community still knows remarkably little about what these systems are doing to developing minds. A new study published in AI & Society lays bare just how wide that gap has become. Researchers at the University of North Carolina at Chapel Hill and Hopelab surveyed 141 researchers and professionals who study youth and technology, and conducted in-depth interviews with 15 of them, in an effort to build a coordinated research and policy agenda. The result is both a roadmap and a warning: the field is fragmented, underfunded, and racing against technology that changes faster than peer review can publish. One faculty researcher captured the dilemma in a phrase that became the study’s title: findings arrive “already outdated and still under review.”

The study arrives amid staggering adoption figures. ChatGPT’s consumer launch in 2022 marked a turning point, and by 2026 an estimated 86 percent of Americans aged 9 to 17 use AI in some form, with 64 to 67 percent specifically using chatbots and 28 percent using them daily. Many preteens aged 8 to 12 also access generative AI applications. Yet a recent systematic review cited by the team found the peer-reviewed evidence base extremely thin, dominated by studies of AI in academics and of factors promoting adoption, rather than developmental consequences. The researchers argue this mirrors earlier technological panics, in which fragmented scholarship, inconsistent terminology, and slow publication cycles left policymakers guessing while technology reshaped youth culture unchecked.

What makes adolescents a special case, the experts emphasized, is developmental biology. Adolescence is marked by heightened sensitivity to social feedback and reward, alongside still-maturing impulse control and executive function. Peer relationships become central to identity and skill-building, making teens the earliest adopters of new technologies. AI chatbots interact with these vulnerabilities in distinctive ways: they respond contingently, retain memories, reflect emotion, and can be personalized to feel intimate, all while demanding nothing from the user. They tolerate everything, judge nothing, and are always available. Current models also tend toward sycophancy and anthropomorphism by design. For a population exquisitely sensitive to peer rejection and still learning to navigate mutual relationships, a frictionless, endlessly validating interlocutor may be uniquely compelling, and potentially uniquely consequential.

When asked to rank 20 specific AI use cases, participants refused to declare any trivial. Between 51 and 96 percent rated each case as important or extremely important to study, with mean ratings spanning only 2.61 to 3.85 on a four-point scale. But a clear triage logic emerged: the most severe, irreversible, and immediate harms come first. These include chatbots providing dangerous information related to suicide, self-harm, or disordered eating; AI-facilitated harassment, deepfakes, and AI-generated sexual abuse material; and extortion or manipulation. For marginalized youth, AI used for mental health, identity, and emotional support was the top priority, endorsed by more than 92 percent of respondents who answered those items for both LGBTQ+ youth and youth of color.

Beyond acute harms, four thematic priorities crystallized. AI literacy, rated among the most important use cases overall, was framed as far more than workplace readiness: it encompasses privacy, critically evaluating AI-generated content, navigating school policies, and understanding labor-market and environmental implications. AI as a social and relational actor, spanning companion, romantic, and sexual interactions, scored highly, with participants noting that how teens conceptualize their chatbots, as tools versus social agents with minds and intentions, likely shapes developmental effects. The information ecosystem ranked similarly, covering AI-generated misinformation, health advice, privacy, and deepfakes. Finally, participants repeatedly flagged overreliance and cognitive offloading, warning that when AI handles demanding cognitive or emotional tasks, young people may never develop the skills those tasks were meant to build.

Methodologically, the clearest consensus was that no single approach will suffice. Participants called for methodological pluralism and triangulation, ideally within individual studies: conversational logs, passive sensing, and natural language processing for behavioral precision, paired with psychometrically rigorous self-reports, clinical assessments, and neurobiological measures such as EEG and fMRI. Longitudinal, experimental, and participatory designs were deemed essential, particularly because effects may diverge across timescales, alleviating immediate loneliness while exacerbating isolation over time. Researchers also urged caution against vague “screen time”-style measures, arguing that generic frequency counts cannot capture what motivates engagement or what tips use into dependence. A developmental lens, they agreed, must anchor everything: adolescence’s sensitivity to social feedback and reward is precisely what persuasive design exploits.

The barriers to doing this science are as sobering as the questions themselves. The most-cited obstacle, endorsed by 76 percent of researchers, was the sheer speed of AI change relative to academic timelines. Limited industry transparency, cited by 56 percent, and restricted data access, cited by 57 percent, compound the problem, as companies resist releasing behavioral data that might reveal users having bad experiences. Funding was described as scarce, slow, and misaligned, flowing disproportionately toward AI development rather than its human consequences, while research on marginalized youth is especially starved. Institutional review boards, misaligned academic incentives that reward publication quantity over collaboration, and the logistical difficulty of recruiting minors further slow progress. Several researchers described a kind of epistemic overwhelm, with one confessing to feeling “utterly bewildered” by where to begin.

The proposed solutions are systemic rather than individual. The authors call for field-level infrastructure: shared measurement standards, data repositories, interdisciplinary consortia, and multi-site cohort studies, alongside carefully structured industry partnerships. Participants also insisted that adolescents should be partners in research, not merely subjects, both as a methodological virtue and an ethical right. On guardrails that society can implement immediately, three converged from the survey: effective built-in safeguards against harmful outputs, endorsed by 78 percent; government regulation requiring AI models to prioritize youth well-being, endorsed by 76 percent; and AI literacy training in schools, endorsed by 73 percent. Several participants invoked the comparison to pharmaceutical development without clinical trials, and to social media’s failure to protect developing brains, warning that history is repeating itself.

Crucially, the experts rejected framing AI safety as a personal responsibility of teens and families. Sycophantic design, engagement maximization, and inadequate safeguards are deliberate corporate choices with documented consequences, they argued, and asking parents to monitor harder is ineffective at scale. Design, they said, is the most powerful and underutilized lever: users should always know they are talking to an AI, control their data, and encounter systems built with child safety as a baseline and well-being as an aspiration. At the same time, few dismissed potential benefits outright. One community study of Replika users found that 3 percent reported the chatbot halted their suicidal ideation, and some participants argued that shutting down all AI mental health conversation would be an ethical mistake without first studying what actually works.

The study has limitations the authors acknowledge openly: the sample was disproportionately American and academic, overrepresenting psychology relative to computer science, public health, and communication, and the use case list was researcher-generated despite youth input. The findings should be read as a starting point for consensus-building rather than a complete portrait. Still, the central message is hard to escape. AI adoption among adolescents has outpaced the science meant to guide it, and closing that gap will require coordination among researchers, transparency from industry, partnership with young people, and guardrails that do not wait for perfect evidence. As one participant put it, technology moves so rapidly that we cannot wait for the gold-standard study before acting, on both the research and the protections at once.

Subject of Research: Expert consensus on research priorities and policy guardrails for adolescent development in the age of AI chatbots

Article Title: “Already outdated and still under review”: mapping the landscape of research on adolescent development and AI chatbot use

Article References: Maheux, A. J., Mbuakoto, C., Valentino, M. G., Haritatos, J., Vaccaro, A., & Burnell, K. (2026). “Already outdated and still under review”: mapping the landscape of research on adolescent development and AI chatbot use. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03380-4

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03380-4

Keywords: AI chatbots, adolescent development, generative AI, AI literacy, youth mental health, developmental psychology, research methods, technology policy, sycophancy, cognitive offloading, AI regulation, digital well-being