Youth wellbeing in the age of AI: a developmental framework for risk and safety

Artificial intelligence (AI) systems are rapidly becoming embedded in how young people learn, communicate, and seek support. Yet current approaches to AI safety remain largely adult-centric, focusing on the prevention of explicit harms such as sexual content, violence, or self-harm, while overlooking the relational and psychological patterns that unfold through repeated interaction. Drawing on developmental science, this commentary argues that adolescent safety cannot be adequately assessed without considering the cognitive, emotional, and social processes that characterize this stage of life. Adolescents—whose identity formation, emotion-regulation capacities, and autonomy are still maturing—face distinct vulnerabilities when interacting with conversational agents. Existing safety frameworks and testing protocols, however, are not designed to detect cumulative risks such as dependency, maladaptive reassurance-seeking, distorted relationship expectations, or erosion of agency. This commentary proposes a developmental framework for AI safety centered on: (1) age-appropriate personalization; (2) developmentally tuned risk assessment; (3) longitudinal wellbeing metrics; and (4) participatory co-design. It further argues that independent, privacy-preserving access to interaction-level data is essential for evaluating developmental impact and enabling continuous oversight. Ensuring that AI systems are safe and beneficial for youth will require integrating developmental expertise throughout model design, evaluation, and governance. Over the past decade, technology and social media have become integral to young people’s daily lives. This has led to extensive debate about their impact on youth mental health. Yet never before has this issue commanded as much public and policy attention as it does with the growing ubiquity of artificial intelligence (AI). Adolescence represents a uniquely sensitive stage of human development—marked by heightened neural plasticity, evolving emotion-regulation capacities, identity formation, and deep attunement to social and peer contexts [ 1 , 2 , 3 , 4 , 5 ]. Against this developmental backdrop, the rapid rise of AI presents both unprecedented opportunities and significant risks for young people’s wellbeing. Recent tragedies involving adolescent suicides linked to AI companions have compelled governments and technology companies to confront these risks. In the United States, California enacted Senate Bill 53—the first law to establish transparency and safety obligations for frontier AI models. The legislation requires consumer conversational agents (i.e. AI chatbots) to implement youth-safety measures, such as periodic disclosures that the agent is not human and crisis-response hand-offs. At the federal level, the U.S. Surgeon General has warned that digital tools have not been proven safe for youth and called for stronger independent evidence [ 6 ], while Congress has reintroduced the Kids Online Safety Act to impose a duty of care on platforms serving minors. International policy developments reflect a similar trajectory. The European Union’s AI Act classifies AI systems that interact with children as high-risk, subjecting them to mandatory conformity assessments, transparency obligations, and human oversight requirements. The United Kingdom’s Age Appropriate Design Code (AADC) takes a complementary approach, requiring online services likely to be accessed by children to conduct Data Protection Impact Assessments and to default to the highest privacy settings for young users. At the international level, UNICEF’s Policy Guidance on AI for Children (2021; updated 2025) outlines nine requirements for child-centered AI systems, drawing on the Convention on the Rights of the Child to establish a normative baseline that AI governance frameworks have only begun to operationalize. These domestic and international policy steps represent meaningful progress but will likely fall short until they approach youth safety through a developmental lens. Children and adolescents are not simply smaller adults; their self-concept, attachment styles, capacity for self-regulation, risk appraisal, and autonomy are still maturing [ 1 , 2 , 3 , 4 , 5 ]. Developmental sensitivities also vary widely by age [ 7 – 8 ]. Understanding and effectively mitigating AI-related risks therefore requires embedding developmental science directly into the design, evaluation, and governance of AI systems.