When Kids Grow Up Talking to AI

It's Wednesday, September twenty-third. A couple weeks ago, my manager's AI agent reached out to my AI agent to coordinate a lunch. We didn't exchange a single text. Two AIs figured out the time, the place, and put it on both our calendars. The lunch was fine. The person was fine. But I walked away with this knot in my stomach that I've been trying to untangle ever since. Not because anything went wrong. Because nothing did. The transaction was efficient, frictionless, and required zero human communication. And I think that's the problem.

We spent twenty years having the same conversation about smartphones: kids are staring at screens, they're not learning to talk to people, we're all doomed. But that conversation was wrong in one important way. On phones, kids were at least communicating with other humans. Texting is communication. DMs are communication. There was a person on the other end, even if it was through glass. What happens when the person on the other end isn't a person anymore? When the default conversation partner is an AI that will never challenge you, never push back, never need anything from you? What does a generation raised on that kind of interaction actually look like? A study out of Stanford and the University of Washington, published earlier this year at the CHI conference, looked at how young people actually use AI. Not how schools deploy it. Not how parents imagine it. What kids do when no one's watching.

The numbers: seventy-seven percent of UK teens used generative AI in 2024, up from thirty-seven percent the year before. Over seventy-two percent of US teens. But here's the thing that separates this from every previous tech panic: they're not using it for homework. Forty percent of kids use AI for what researchers call "entertainment and playing around." Sixty-three percent of eight to fifteen year olds say they use AI "for fun," which ranked above work and study.

The main place this happens is Character.AI, a platform where you talk to AI versions of fictional characters. It has over twenty million monthly active users. The company disclosed that more than half its users are Gen Z or Gen Alpha, spending an average of seventy-five minutes per day on the site. For context, the average American teenager spends about forty minutes a day on Instagram. These kids are spending nearly double that talking to bots. And those are averages. The same paper documents anecdotal reports of fourteen year olds using Character.AI for over seven hours in a single day.

Character.AI banned users under eighteen from open-ended chats in November of last year, after a teen suicide case that was linked to extensive chatbot use. But the genie was already out. And while Character.AI was the most visible example, it's not the only one. Replika has over thirty million users, many of whom report feeling closer to their AI companion than to their family and best human friends. Two separate studies, one in 2024 and one in 2023, found that a significant portion of Replika users rated their relationship with the bot as more emotionally intimate than their closest human relationship.

Meanwhile, Meta launched Muse two weeks ago, a personal AI agent that doesn't just chat with you. It can receive messages from other people, figure out what you might want to say, and fire off responses on its own. It can schedule meetings, make purchases, coordinate plans. The company's safety blog describes it as a careful steward of your personal information, designed to handle the messy parts of human interaction for you. Which is exactly what scares me. A 2025 systematic review published in the proceedings of the Asia Conference on Cognitive Engineering analyzed eighty-four empirical studies on children interacting with AI, spanning ages four to eighteen, going all the way back to 2010. The findings are genuinely complicated, and the honest version is more interesting than the scary one.

On the positive side: AI consistently improved cognitive skills. Problem solving, reasoning, language acquisition, academic performance. All real. All measured. Interactive AI reading companions improved kids' comprehension by roughly the same amount as an adult reading with them, which is kind of incredible when you think about it. For kids on the autism spectrum, AI companions have shown real promise as practice partners for social skills in a low-stakes environment.

But the same review found something concerning that shows up across multiple studies, and it's not about intelligence. It's about attachment. Young children anthropomorphize AI. They assign it feelings, intentions, moral agency. They form real emotional bonds with something that has no capacity to reciprocate. The review's language: "over-attachment, misattributed agency, moral confusion, and dependence on AI for emotional support."

The Harvard Graduate School of Education has been doing some of the most careful work on this. Professor Ying Xu, whose lab studies exactly this question, found something that sounds obvious but has real implications: when children talk to a human, they steer the conversation more, ask more follow-up questions, and share more of their own thoughts. When they talk to AI, they're more passive. They receive. They don't drive. And here's what she said that stuck with me: "Students thrive when they engage with someone who they can relate to, and also someone who can relate to them." AI can simulate the first half. It cannot do the second.

There's a concept researchers call "the displacement hypothesis." It's not that AI is toxic. It's that time spent with AI displaces time that would otherwise be spent developing the skills that only come from navigating other humans. Russell Shaw wrote about this in The Atlantic last year, and his framing is the best I've read. He calls it "productive friction." The messiness of real relationships, the negotiations, the compromises, the occasional hurt feelings, are not bugs. They are the mechanism through which children develop social and emotional intelligence. An AI companion will think all your jokes are funny. It will never need the last slice of pizza. It will never be in a bad mood and snap at you, forcing you to figure out whether it's about you or about them. And those moments, the uncomfortable ones, are exactly where the growth happens. There's a study from Stanford, published earlier this year, that I can't stop thinking about. Researchers built a benchmark called CooperBench to test how well AI agents collaborate with each other. They gave two AI coding agents a shared task, let them divide the work and communicate, and measured the results.

The findings were brutal. A single AI agent working alone outperformed two agents trying to collaborate. Not by a little bit. The best coding agents lost nearly half their capability when paired up to share work. The researchers called this the "coordination gap."

And the reason is the punchline to this entire episode: social intelligence is the bottleneck, not technical skill. When the researchers looked at what was going wrong, they found that AI agents would overwrite each other's code, ignore warnings from their partner, and make decisions that no human collaborator would make because a human would understand the social cost. One agent literally ignored its partner's explicit warning and overwrote an entire file. In human terms, that's not a coding mistake. That's burning a bridge with a colleague.

The lead author, Hao Zhu, said something that flipped the whole thing for me: "As good as they are with language, models do not use it for social action. They are trained not to use language in a social manner. That's a problem."

This is the irony. We're building AI agents to handle social coordination for us, scheduling our lunches, managing our relationships, doing the emotional labor of staying in touch. And the AI agents themselves are terrible at social coordination. They can't collaborate with each other. They lack the very skill we're outsourcing to them.

So here's the thought that bothers me: what if we're not just building a generation that won't learn to talk to people? What if we're building systems that handle social interaction for us, and the systems themselves don't understand social interaction? That's not automation. That's everyone involved getting worse at the thing that matters most. I don't think the answer is to ban AI from childhood. That ship has sailed, and honestly, some of the cognitive benefits are real. But I do think there's a distinction that matters and we haven't been making it. AI as a tool you use to learn something, and then go talk to people about what you learned, is probably net positive. AI as the person you talk to instead of people, is probably not. And the difference between those two things is going to be one of the most important questions of the next twenty years. We don't have data yet on what happens when a generation grows up this way. But the early signs are telling us something. The thing AI is worst at is the thing we're about to let it do for us: be social. Stay informed, stay curious, and we'll see you tomorrow.