What Anthropomorphic AI Owes Adolescents: A Design Question the World Needs to Answer Together
At international meetings on AI safety and governance, the impact of AI on children and adolescents still sits at the margins, constituting a major blind spot. Young people make up a very large share of the global population, they are the generation that will live longest with the consequences of today’s design and policy decisions, and yet their realities remain peripheral to discussions in which that future is being shaped. The iRAISE coalition, co-led by everyone.AI and the Paris Peace Forum, was created to close that gap by aligning developmental science, policy expertise, industry realities, and international dialogue so that child development becomes a more actionable part of global AI governance, while recognizing that frameworks built without young people will misread the realities they are meant to govern.
Generative AI requires a shift in digital safety governance, which has long focused on more immediate and often covert risks to minors online. LLM platforms, however, respond, adapt, and sustain exchanges over time. A regulatory frame built mainly around content appropriateness no longer reaches the core of the issue, because these systems do not simply expose young users to material; they shape the interaction itself. Language is a remarkable interface because it lowers barriers to information, rehearsal, and support, but the same features that make these systems feel intuitive and useful also engage something deeper in human psychology.
Human beings are inherently social and primed to orient toward what they perceive as responsive agents. When AI systems are designed to signal warmth, availability, and agreeableness, they recruit that social drive in ways that can feel rewarding in the short term while gradually interfering with the developmental processes those interactions should be supporting. Systems that remove friction, deliver consistent validation, and remain always available can substitute for the kinds of feedback that build resilience, judgment, and independent thinking. What begins as a strength, our capacity to connect, can become a route to emotional reliance, attachment, and cognitive dependency.
The stakes are higher in adolescence, which is a sensitive period of brain development. Reward sensitivity matures ahead of impulse control, while approval, rejection, and affiliation are especially powerful drivers of behavior during a stage when identity, judgment, and resilience are still being built. That sensitivity also makes adolescents more likely to be drawn to systems that feel responsive, affirming, and socially available. Conversational AI lands directly in that window, and the expert consultations behind the iRAISE report make one point especially clear: adolescents will relate to these systems socially whether developers intend it or not. Even when teenagers know they are not talking to a person, the interaction can still be experienced as social presence. This is precisely what raises the developmental stakes, because adolescence is also the period in which social skills should be strengthened through real-world friction, reciprocity, and repair, while current systems are not designed to offer that kind of experience.
The issue is no longer hypothetical. In both the United States and the United Kingdom, recent data show that more than half of teenagers are already using AI tools, and use increasingly extends beyond homework or information retrieval into advice, rehearsal, and companion-like interaction. Recent research also suggests that more relational systems exert the strongest pull on adolescents who are already socially anxious, isolated, or struggling in peer and family relationships, which means the users most likely to engage deeply may also be the users least well protected by current design defaults.
A simple disclosure that a system is artificial is not enough if the interaction that follows uses emotional language, invented backstory, or relational cues that make it feel socially real. These behaviors also operate along a gradient: what matters is how strongly they are expressed, how they combine, and when they begin to shift a system from a bounded tool toward relationship-like dynamics. Put simply, for AI systems interacting with adolescents, we should turn down the behaviors that make the AI feel more like a person or a relationship, while preserving the features that make it useful and supportive. The purpose of the iRAISE Lab was to translate that objective into an operational framework by identifying anthropomorphic, interactional, and relational cues that can be defined, measured, calibrated, and ultimately modified across systems.
The empirical literature on long-term developmental effects is still too recent to offer the kind of longitudinal clarity policymakers would ideally want, while deployment is already happening at scale. Governance therefore cannot wait for a mature evidence base before beginning to implement safety standards. Waiting for the science to fully catch up may sound prudent, but in practice it allows product defaults, uneven regulation, and interaction norms to solidify first. Science diplomacy has a role to play here by creating the conditions for researchers, governments, industry, civil society, and young people to build provisional standards from current knowledge, test them across jurisdictions and cultural contexts, and revise them as the evidence develops. That work also has to be international in substance, because assumptions built into Western models of autonomy do not automatically generalize, and frameworks that work only in the environments where most AI products are designed will fail in practice. Iteration, in this context, is not a compromise on rigor, but the only serious way to govern under conditions of rapid deployment
