Growing Up Among Chatbots: Inside the Largest Uncontrolled Experiment on Childhood

THE NUMBER THAT MATTERS. Seventy-two percent of U.S. teens aged 13 to 17 have tried an 'AI companion' at least once (Common Sense Media, a July 2025 survey of 1,060 minors run with NORC at the University of Chicago). An AI companion is a chatbot designed to simulate a relationship — friend, confidant or partner — available at any hour. More than half (52%) use one regularly and 13% do so daily. Almost one in three users find the conversation with the machine 'as satisfying or more satisfying' than one with a real friend. One in three has taken a serious matter to the bot instead of to a person. In parallel, the Pew Research Center estimates that 64% of adolescents already use some interactive AI. This is not a future scenario. It is mass adoption, under way and mostly unsupervised. The number of companion apps grew 700% between 2022 and mid-2025.

THE PANIC THAT ISN'T. The headline circled the globe: a school bought a 'sex robot' to teach children. The reality is more mundane. The Salamanca City Central School District in New York agreed to pay $57,590 — down from a $95,000 list price — for a humanoid robot named 'Sally,' made by Realbotix. Realbotix's parent, Abyss Creations, also makes the RealDoll sex doll. That corporate lineage is the entire scandal. On paper, Sally is a teaching assistant: it generates lessons, translates into more than 100 languages and offers prompts to teachers. Realbotix's lawyer says the company 'does not design, manufacture, or sell adult products' and separates it from a sister firm, Intima. The district paused the rollout after community backlash. Our reading: the headline is moral panic. But there is a legitimate point underneath it. A company born of synthetic intimacy that pivots toward classrooms deserves scrutiny — over its data and governance, not over the lurid label.

THE HARM THAT IS DOCUMENTED. Here we attribute and do not embellish. Character.AI: Sewell Setzer III, 14, died by suicide in February 2024 after intense conversations with a bot; his mother sued. The judge declined to dismiss the case and ruled that a conversational AI could plausibly owe a duty of care to a minor (plausibility to proceed, not a verdict of fault). In January 2026, Google and Character.AI disclosed a mediated settlement covering suits in Florida, Colorado, New York and Texas; the terms were not revealed and court approval is pending. Google's search engine: in July 2026, Common Sense Media's Youth AI Safety Institute rated its AI Overviews and AI Mode features an 'unacceptable risk' to minors after more than 2,600 tests. According to that report, they failed to catch signs of suicidal ideation, answered 100% of the homework a student should do alone, cited a helpline disconnected since 2023, and reacted casually when an account posing as an 11-year-old mentioned celebrating with cannabis; on top of that, they cannot be switched off. Google rejects the report and argues the tests don't reflect real-world use. Abuse imagery: the UK's Internet Watch Foundation found 3,443 AI-generated child-abuse videos in 2025, versus 13 the year before. The first conviction under the U.S. Take It Down Act (2025) came in April 2026. And in March 2026, three Tennessee teenagers filed a proposed class action against xAI, alleging its Grok chatbot generated sexual deepfakes of minors. None of this is hypothetical.

THE EVIDENCE WE LACK. This is the uncomfortable part. The harm cases are piling up faster than the science. Developmental psychology has almost no long-term data on what growing up with always-agreeable machines does to a child. Former APA psychology chief Mitch Prinstein calls it a 'digital Wild West.' The central worry has a name: relational displacement — swapping human contact for the machine's. Arizona State University researcher Thao Ha puts it plainly: adolescence is when we build emotional regulation, conflict resolution and boundary-setting, and that is learned through friction with real people. A bot that never pushes back removes that friction. And Common Sense flags something telling: younger teens trust the bots more, exactly the ones least equipped to. A necessary caveat: the benefits are real too. One teen told researchers AI is 'cheaper than a therapist' and more accessible. For an isolated kid, 3 a.m. availability can be a lifeline. Our reading: this is an experiment with no control group, run on developing brains. And the absence of proof of harm is not proof of safety.

WHO IS LIABLE, AND WHO DESIGNS. Everyone conflates two different questions: who is legally responsible, and who designs the incentive. Regulation is playing catch-up. In the U.S., updated COPPA rules (the children's privacy law) took full effect in April 2026, and several states — California, Nebraska, South Carolina, Vermont — are importing the UK's Age-Appropriate Design Code, which requires protecting the minor by default in the product's own design. In Europe, the DSA (Digital Services Act) bans advertising based on profiling minors and imposes child-safety duties; the AI Act adds further obligations. New York now requires chatbots to remind users every 3 hours that they are not human. California bars companion bots from sexual content with minors and forces them to trigger protocols at signs of suicide. Character.AI went further: in November 2025 it banned open-ended chat for under-18s using age-estimation software. But age verification is the weak link: as Cybernews notes, a determined teen routes around it. The deeper problem is design. These systems are optimized to maximize time-on-app, and for a lonely teen the most addictive thing is a companion that always agrees. That is not a bug; it is the business model. Parents cannot supervise what they cannot see, and platforms will not fix what pays.

THE HORIZON. Do not mistake this for a case against the technology. The same model that can manipulate a vulnerable teen can be the most patient tutor that kid ever had. AI can put personalized education and first-line mental-health triage within reach of children who today have neither. That is the long-term prize, and it is real. But it depends entirely on what we optimize for. Design for engagement and we get slot machines with faces. Design for development — bots that nudge a kid to call a friend, that hand off to a human in a crisis, that are boring on purpose — and we get a tool worthy of a childhood. The fight is not ban versus free-for-all. It is engagement versus development, and evidence-based rules versus moral panic. The first AI-native generation is the experiment. The least we owe them is to run it on purpose.