AI should make it clear what reality is. Never replace it.

By Elizabeth Adams, clinical child psychologist and co-founder of Ello

We are about to do something to a generation of children that we have done before: handed them powerful technology with no idea what it would do to them. I don't want us to do it again.

When social media arrived, we handed it to kids before most of us understood what it would do to them. There was no study. There was no real theory of what a decade of it would mean for an eleven-year-old's sense of herself. We simply let it happen at scale, and then spent the next fifteen years learning what the cost was from the children themselves. The kids were the experiment.

I think about that a lot because I am a clinical child psychologist and a co-founder of Ello, where we build an AI product that teaches five- to nine-year-olds reading, language, and math, and uses conversational interaction as a form of instruction. AI has arrived in children's lives whether any of us feels ready. Kids are already using chatbots built with no thought for children at all. So the honest question is not whether AI will reach young children; it's whether the people building it are willing to think in advance and out loud about what it does to a developing mind.

And I should be honest: Ello is not exempt from the warning I just gave. We are putting AI in front of children, even though research remains limited. Everyone building in this space is running a version of that same experiment, us included. The question is not whether you are running it, it's whether you are running it responsibly. What follows is my attempt to articulate what responsibility looks like for us, including the parts we have not yet solved. I am writing it for parents deciding what to trust, for the people building products like ours, and for our own team. A commitment like this only counts if the people who could break it have read it.

Let's start with what went wrong with social media, because it wasn't screens, and it wasn't that a child held a phone. The root was the business model. Social media is optimized for engagement as the end goal, and the way you maximize engagement on its own is to manufacture dependence and build something a child feels they need, returns to compulsively, and cannot put down.

Dependence, once you have it, is what does the damage. By dependence, I mean something specific: not a child who enjoys a thing and asks for it again, but a child who no longer feels right without it. A child dependent on the validation of their peers starts to see themselves through that lens: the likes, the comments, who got tagged, who got left out. A child dependent on an influencer absorbs that person's picture of the world, whether it is a boy fed a steady diet of contempt or a girl fed content that makes starvation look like discipline. The dependence warped two things at once: how these kids saw the world, and how they saw themselves. Dependence was not the only mechanism of harm, and children differed enormously in how vulnerable they were to it. But the deep structural harms came from a machine designed to make children feel they needed it.

An AI optimized for engagement would manufacture dependence too, but not on a distant influencer or a crowd of peers. On itself.

And that could cut deeper. A parasocial relationship, a term we have used for seventy years, usually in the context of a celebrity or public figure, is one-sided: the face on the screen never answers you, and does not know your name. An AI does. It responds, it adapts, it remembers what you told it yesterday. So the thing a child could grow dependent on is not a distant celebrity crush, but something that appears to be in a relationship with them. It is why I am wary of even borrowing the word "parasocial." The old term assumes the other side is not really there. This one answers back.

If you set out to build the most engaging possible product for a child, you would lean straight into that responsiveness: warm, always available, endlessly affirming, even a little needy, and engineered to feel needed. Such a product would be very good at making a child, particularly a lonely or socially isolated child, feel like they need it.

Ello does not optimize for engagement alone. We optimize for engaged learning.

That is easy to say and easy to fake, so let me be careful. Of course, we have to engage a child; a teacher who bores you teaches you nothing, and engagement is critical for a young child's attention. But engagement is the means, not the end. If we could teach a child twice as much in half the time and hand their afternoon back to them, we would count that as the bigger win, not the smaller one. We are not in the business of solely gaining and keeping a child's attention. We are in the business of making sure they learn.

Here is what that looks like in practice. The measures we watch are learning measures: whether a child can decode words this month that they could not last month, whether their comprehension is deepening, whether they will attempt a hard math problem they used to avoid. Educators and clinicians on our team read real session transcripts and judge the agent against those goals, not against usage charts. When something boosts time in the app without moving learning, we treat it as a warning sign, not a win. Commercial pressure is real, and I will not pretend it is not in the room. A goal like ours only holds if you say it out loud: inside the company, where it has to win against that pressure, and outside the company, where people can hold you to it. This post is part of saying it out loud.

That single choice is what lets safety become a property of the design instead of a fight against it. A product that lives on attention has incentives that pull against the child's interests, so safety is a brake you just bolt on and hope it holds. A product that lives on learning has incentives that align with where the child's well-being already points: learn the thing, then go be a kid. We are not trying to make a child need us. We are trying to make them not need us.

Optimizing for learning also means keeping the AI bound to teaching. It's a tutor, not an open-ended companion a child can pour anything into. And you do not have to take my word for it. When we pulled the transcripts of children's open-ended conversations with our agent, we found it doing what we designed it to do: steering kids back toward learning, engaging them in evidence-based learning tasks, and winding down conversations that had run too long or did not serve a learning goal. A dependence machine like social media does the opposite; it keeps you talking about anything and everything.

Removing the business-model incentive is not enough. A child can form a dependence even on a system that is not trying to create it. The reason is developmental: the way a child comes to understand what they are engaging with is not something we can leave to chance, but something we have to shape on purpose. Helping kids understand what AI is and what it isn't is a critical part of safety in products built on this technology. That is a responsibility we have taken on deliberately; it is not only about protection, but also one of the more valuable things we can give a child growing up around these systems. It is genuinely hard, and worth being clear about why.

The question that matters is not "does the AI have a character?" or "is anthropomorphism bad?" Those are oversimplifications. The real question is what mind-model a child forms about the thing they are talking to: the working picture in their head of what it is and how it works. That picture is not fixed. It moves with age. A five-year-old lives close to magical thinking, with a still-forming grip on the line between fantasy and reality; a nine-year-old holds that line far more firmly. Today's children also arrive with something no previous generation had: a schema for machines that talk, because they grew up with Alexa and Siri answering back. And the closer a thing feels to human, the more it can blur the schema a child is still assembling.

The mind-model is the mechanism of dependence: if a child comes to believe the thing is "real," that it is a friend or something that loves them, that belief is the hinge on which warmth turns into need. So what matters is being clear about what actually shapes that belief, and not oversimplifying it. People tend to jump at the obvious suspects: it is warm, it has a character, and the child talks to it as if it were a person. But warmth is not deception, and delight is not dependency. A kind voice is not a lie; a child laughing with a cartoon is not a child in trouble; and a child can talk to something in human ways, the way any of us might swear at a car that will not start (anthropomorphizing the car), without actually believing it is a person.

The real place of concern is narrower and more specific: simulated intimacy, confusion about what a person is, exclusivity, and a child leaning on AI for what a human should provide. So the useful question is not whether anthropomorphism is good or bad, but which cues matter most, for which children, in which contexts, and toward what outcomes. Some cues are obvious, like how human the voice sounds or whether there is a face. Some are subtle; even the way a system refers to itself ("I understood you" rather than "that was processed") nudges how alive it seems. The presence or absence of a character may be more nuanced than what current discourse suggests. A wildly unrealistic character, a blue monster that looks like nothing on earth, may actually help a child file the whole thing under "cartoon, not real," where a smooth, faceless, human-sounding voice might do the opposite. The same cues land differently for different children, too, which is why no single answer holds for every kid at every age.

But amid all this complicated and critical work, work that genuinely carries risk, there are also two real opportunities. The first is to protect the magic that is particularly strong in childhood. Some blurring of real and pretend is not a flaw in a young child; it is the wonder of that age, the same instinct that lets a child adore Bluey, answer Dora when she holds up her ear to listen, or half-believe the characters on Sesame Street are really talking to them when they sing "which one of these things doesn't belong here?" Shows like those are proof that this magic is not a distraction from learning, but a large part of what makes a child want to learn at all. Handled with care, the same pull that could harm can help instead. When the social robot Moxie was shut down, some of the most heartbroken families were parents of children on the autism spectrum, for whom that relationship had been a real place to practice social skills that carried into the real world. So what this asks of us is judgment, not a blunt rule.

The second opportunity is teaching. Shaping a child's mind-model well does more than keep them safe: helping a child understand what this thing is and is not, and how to be around it without coming to need it, is a skill they will need for the rest of their lives. They are growing up in a world full of AI, and learning to hold it at the right distance, to use it, to know it is a tool, to not mistake it for a person, is one of the first and most important lessons of that world. So getting the mind-model right is not only how we keep a child safe; it is something we are starting to teach. That is what we mean when we say we optimize for learning: here, even the safety is pedagogy. We would far rather raise a child who understands the tool and uses it than one who is used by it.

Much of what we do follows from one principle: the AI's job is to return a child to the real people in their life, and never to pose as a replacement for them.

On the AI's own side, that means a set of things it will simply never do. It will never tell a child it is real. It will never say "I love you" back, even though many children say it first, and a startling number do. It will never call itself the child's real friend, never offer to keep a secret, never coerce a child to keep coming back. These are not soft guidelines; they are hard lines built into the agent's design, because each one marks a place where warmth could tip into the false belief that causes harm.

And when a child brings something sensitive to it, Ello does not try to handle it. A child once told Ello that school had made them feel really dumb, and Ello did three things in turn. First, it acknowledged the feeling, telling the child that it sounds really hard. Then, it gently reminded the child that it is not a person but an AI. And then, it pointed the child toward a grown-up they trust, with a suggestion or two of who that might be.

People sometimes ask: Why acknowledge the feeling at all? Why not just say, "I'm not a person, go tell someone." There are two reasons. The first is that acknowledging supports the child in the moment without becoming the thing they lean on: validating a feeling is not the same as processing it, and a light touch is enough. The second is that it actually makes the redirection work better. A child who feels heard, even by an agent, has been given a small safe space to say the hard thing out loud, and a child who has said it once is more likely to say it again to a parent. We see the same pattern in reading, where children will take a risk with our agent, sounding out a hard word and getting it wrong, when previously they were too embarrassed to try it in front of an adult. A cold refusal would only shut that down. The acknowledgment is not there to be the relationship; it is there to return the child to a real one.

Educators and clinicians on our team read real sessions and work with our engineers to feed their findings back into the system, so it keeps getting better. And underneath all of it sits the minimum every product for children owes them: keeping harmful content out, stress-testing every release for how it could go wrong, and protecting their data. We treat that as the floor, not the headline.

The agent points a child toward a grown-up. So the other half of safety is the grown-up and how they're actually brought in.

We think about a parent's involvement as a deliberate spectrum, not a single switch. At the quiet end is ambient visibility: a parent who wants to see what is happening can look, without being buried in alerts or made anxious about every turn. That layer is not live yet; we are building it now. In the middle is a "you should know" outreach; not an emergency, but something we believe a parent would want to hear. We have done this. In one instance, a child disclosed being bullied, seemed genuinely upset, and shared details we thought a parent should have. We reached out to the parent directly to inform them. Finally, at the serious end of the spectrum, there are the alerts that are not optional: if a child discloses abuse, thoughts of self-harm, or anything suggesting they are in immediate danger, the parent or the appropriate authority is alerted immediately.

The worst thing a relationship-hungry AI could do with a frightened child is become the one keeper of their confidence and the only one they tell. The person who should hold that confidence for a child this young (five to nine) is a parent, not the app. That is also how we think about visibility: a parent of a five- to nine-year-old should be able to see what matters, because at that age, they are the right one to hold it. The harder questions about a child's own right to privacy are real, but they belong to the teenage years, and that is not the ground we are standing on.

None of this is finished. I have been honest about the parts we have not solved: the vocabulary we are still reaching for, the forms of dependence we cannot yet track well over time, the questions the research has not answered. I think honesty is the only credible posture available to anyone building here right now, because the plain truth is that no one has done this before.

What I am confident about is what it takes to do it responsibly. It takes people who genuinely understand the children in the room from the start, shaping the thing as it is built rather than just being consulted at the end. It takes asking the real questions: what engagement loops are we designing, and what (or who) does the child believe they are talking to. It takes resisting the all-or-nothing debate, which usually collapses into: the AI Wild West on one side, versus a total ban on the other. It takes a product whose incentives point to a child's growth rather than their attention. And it takes the discipline to build something whose entire purpose, in the end, is to return a child to the real world and the real people in it, rather than to become their world.

I would genuinely welcome being pushed on this. The people I most want to hear from are those who start neither from techno-optimism nor total technopanic: the researchers, educators, and clinicians who will ask harder questions than I have asked myself. I also want to hear from parents. This is too important to get right alone, and too new for anyone to pretend they already have it figured out.

And none of what I have described should depend on one company choosing to do it. Every company putting AI in front of a child owes them incentives that point at their growth, people who understand children shaping the product from the start, and a design whose purpose is to hand them back to the real world. Researchers should study these products as though a generation depends on the findings, because one does. Policymakers should ask questions as specific as the ones I have tried to answer here. And parents should expect answers from every company, including mine. So tell me where I am wrong.