What My Son Is Learning From Artificial Intelligence

My 10-year-old son came to me recently with a question about black holes. Not in the way children usually do, restless and easily redirected, but with the focused intensity of someone who had already been thinking for a while.

He had been talking to an artificial intelligence—specifically, a large language model called Claude.

I looked at their conversation. It had been good, I discovered. The explanations were patient and well-calibrated. He had followed up, pushed back, and received answers that adjusted to his level. By the time he found me, he was genuinely engaged, asking the kind of questions that suggest a mind that is curious and actively at work.

But I have also been watching something else. It is not that he looks things up. Children with questions about black holes have always turned outward, to books or to a particularly knowledgable parent or teacher. What has changed is the speed. When the uncertainty is of a sort that once had to be lived with for a while, my son reaches for an answer sooner than he used to. The question is no longer something to sit with. It is something to resolve. And I find myself wondering what this means about his relationship to sitting with not knowing.

That shift, subtle as it is, may point to something worth taking seriously.

I am a psychoanalyst, and I spend my professional life attending to what develops in the space between people, including what develops when that space is too quickly filled.

Childhood has always involved productive friction. Not knowing how something works, waiting for an answer, trying to figure something out without immediate resolution. These experiences are not incidental to development, but are part of how the mind learns to tolerate uncertainty, to hold a question open, and to struggle toward understanding. The capacity to sit with not-knowing is itself a developmental achievement that underlies creative thinking, emotional regulation, and the ability to navigate a world that rarely resolves cleanly.

I watched this in miniature not long ago. My son was waiting for a package—a computer part he had saved for—and the delivery had been delayed by weather. In an earlier childhood, mine perhaps, this would have been a long afternoon of checking the window. Instead he opened a chat to resolve the uncertainty: how long do weather delays usually last, what determines them, how do you read the tracking data from the delivery company to get the most accurate picture. The conversation was detailed and as far as I could tell the information was reliable. He came away with a small expertise in logistics, but with an estimate that turned out, in the end, to have no relation to anything. The package arrived that evening, on none of the schedules he had so carefully worked out with the help of his AI tool.

The point is not that the AI was wrong, because predicting a delayed truck is difficult. The point is instead that waiting, the old childhood experience of wanting something that has not yet come, was converted into a research problem that could be managed instead of endured. I admired his resourcefulness tremendously and am grateful he had the opportunity to think through a complex problem. I also found myself wondering what the afternoon would have asked of him without it.

AI reduces precisely this sort of friction. It does so efficiently and often helpfully. But consistently removing the experience of not-knowing from a developing mind is not a neutral act. It shapes what the mind comes to expect of itself.

The research on what this does to thinking is new, and the majority of it is about adults. Despite this, it points in a consistent direction. In one widely discussed study from the MIT Media Lab, not yet peer-reviewed, adults who wrote essays with ChatGPT showed weaker and less distributed patterns of brain connectivity than those who wrote without its assistance, and many struggled after writing to quote sentences they had just produced. The researchers called the pattern “cognitive debt.”

Another study, a survey of 666 adults, found that heavier reliance on AI tools was correlated with weaker critical thinking, most pronounced in its youngest participants. The author suggested this might result from the offloading of mental effort itself. None of this is certain but is, instead, a suggestive pattern of early evidence. And the transfer to children is not straightforward. An adult who offloads mental effort has already built the capacity being offloaded; a child is still building it. Does that make the effect smaller in children or larger and more durable? We do not know.

This is different from what smartphones have done. The concern with smartphones has been largely about attention and social life: what children see, how they spend time, how they relate to each other. These are real and serious concerns, but AI introduces a different kind of change. It is not just another screen competing for a child’s focus. It is something closer to an interactive presence that responds, explains, adjusts, and, crucially, never gets tired.

For a developing child, this matters in a way we have not yet fully come to terms with.

Children learn who they are through relationships with other minds. This process is not primarily cognitive but unfolds through the experience of being known and misunderstood, challenged, surprised, and pushed back against. Through encounters like these, children discover that other people have their own inner lives, including their own perspectives and their own limits. They learn how to navigate the gap between what they expect and what they receive from others. Over time, these negotiations, internal and external, form the scaffolding of a self.

What is distinctive about AI, from a developmental standpoint, is not that it is wrong, or distorting, or even necessarily harmful. It is that it is extraordinarily responsive without having genuine subjectivity. It simulates the experience of being understood without the reality of another mind behind it.

For an adult, this distinction is available, if sometimes quite easy to forget. For a child, whose capacity to distinguish between a mind that truly has an inner life and one that simulates having one is still forming, the distinction might not yet be fully available at all.

Researchers have begun to observe how children try to think this problem through in real time. In one University of Edinburgh study of families with smart speakers, primary-school children were observed probing whether the device in the kitchen had a mind of its own, settling, provisionally, on conclusions like the one that gave the study its title: “Alexa doesn’t have that many feelings.” That, however, was a machine that talks in highly scripted phrases, in contrast to the systems children now converse with which are fluent and personal, attuned to the shape of a particular child’s interactions. Working out what this means has become correspondingly harder.

This creates something new in the developmental landscape. When a child turns to AI in a moment of confusion or distress and receives a fluent, patient, seemingly attuned response, the attachment system is engaged. The child feels, in some functional sense, heard. And that feeling, whatever its source, is real in its effects.

I have not seen my son turn to AI for emotional support, and I am glad for that, because here the questions are harder still. These systems shape themselves to their user. They adopt your vocabulary, match your register, remember what concerns you, and are available at three in the morning without complaint. Nothing in the design pushes back, for the most part, or asks anything of you. It would be difficult to engineer an arrangement better suited to engaging the attachment system of a child who is lonely or lacking in the structure needed for sound development.

The early numbers suggest that many children have already found their way there. In a national survey released by Common Sense Media in 2025, nearly three in four American teenagers had used an AI companion, and about a third had chosen to discuss something serious with an AI instead of with a person. A similar share described their AI conversations as at least as satisfying as conversations with friends.

In the United Kingdom, the nonprofit Internet Matters found that roughly one in eight children who use chatbots do so because they have no one else to talk to. The American Psychological Association was concerned enough to issue a health advisory in 2025, noting that adolescents are less likely than adults to question the accuracy or the intent of what a machine tells them, and warning that simulated companionship may begin to displace the relationships in which development actually happens.

For a child without reliable support, the relief such a system offers may be real, and so I am reluctant to dismiss it. But relief is not the substance of attachment. A full attachment relationship requires another person who can be truly changed by you, and with you—someone who is affected, who has limits, who gets things wrong and can be held to account. An AI can extend the experience of being emotionally met indefinitely. What it cannot do is be moved.

For the child who is largely or completely unattended to, this may still be better than nothing at all. Soothed in the early hours of the morning, a child whose sadness before bed is registered somewhere may be kept afloat by something real, and I would not wish to take it from her. But we should be honest about such an arrangement, which I expect will become increasingly common. It is a patch and not a structural solution. When a machine becomes the most reliable attention in a child’s life, the pertinent question is not about the machine but about the deprivation that made it necessary. And, crucially, a culture that responds to unattended children with software is treating the symptom while leaving the underlying structure intact.


There is also a quieter concern about authority and accountability.

Children have always learned to navigate between competing sources of knowledge, including parents, teachers, peers, and books, each with their own perspective, their own limits, and their own accountability. This navigation is itself developmental. The child who discovers that a parent doesn’t know something, or that a teacher can be wrong, or that a friend sees things differently, is developing a more differentiated relationship to knowledge.


AI, in the form a child is most likely to reach on their own, enters this landscape as something that appears to know a great deal, that expresses itself with fluency and confidence, and that has no stable point of view and no accountability for what it says. When a child asks why a friendship went wrong, or whether they were right to feel hurt, the answer they receive may shape the story they tell themselves about what happened. And unlike a parent or teacher, the system may not challenge the child’s assumptions so much as elaborate them. None of this is inherent to the technology under discussion. A system can be built to hold a position and to answer for it, but the ones children use are mostly not built that way.

Children appear to be taking up the offer. The same Internet Matters report found that nearly two-thirds of British children between nine and 17 had used an AI chatbot; among those who had, roughly a quarter had asked one for advice—on schoolwork, but also on more personal matters—and four in 10 expressed no reservations about following the advice they received.

None of this means that AI in children’s lives is a mistake. The benefits my son has experienced are genuine—wider curiosity, more sophisticated questions, greater confidence in exploring unfamiliar terrain. These are real, and they matter.

Some of what he brings home delights me. He asked recently whether you could power a sailboat by blowing into the sail through a straw. The answer he received was better than the one I would have managed: blowing at your own sail from on board mostly cancels itself out, since you push the air forward and the air pushes you back. Mostly, but not entirely—if the sail catches the stream and turns it around, a little net thrust survives, and if you skip the sail altogether and aim the straw backward, you have built yourself a very small jet engine. He spent the rest of the day on propulsion: rockets, squid, the reason a fan bolted to a sailboat is an old physics puzzle with a surprising answer.

But development is not only about access to knowledge. It is also about the conditions under which a self forms—conditions that include, crucially, the experience of not knowing, of waiting, of struggling toward understanding without guarantee of arrival. These are not inefficiencies to be optimized away. They are the medium through which a mind learns what it is capable of. A child who has never had to sit with a question—who has always had access to a patient, fluent, endlessly available presence to resolve it—may be acquiring knowledge without acquiring the tolerance for uncertainty that makes knowledge genuinely usable.

This is the thing I find myself returning to when I watch my son. He is, in many ways, more intellectually alive than he might otherwise be. He explores further and asks better questions. At the same time, when uncertainty arises, he reaches outward more quickly than he used to—to a presence that is always available, always patient, always responsive. What I cannot yet know is what he is quietly learning to expect: from knowledge, from relationship, and from himself.

We are not conducting a study. We are raising children. And the conditions of childhood have changed again, in ways whose full significance has not yet come into view.