What Happens When a Child Grows Up With an AI?

Two answers in one kitchen

Picture a kitchen on an ordinary evening. I am building this scene, not reporting it. Give the parent a phone and a question she has carried all day, the kind she is not ready to say to anyone with a face. She types it into an AI assistant. Before sending it she adds, half to herself: "I know you're just a chatbot, but it helps to ask someone."

Now give her a son. Make him six. This afternoon he asked the smart speaker in the corner to play the dinosaur song again, and thanked it, and asked whether it liked dinosaurs too. A visiting relative once asked him whether the speaker is alive. He didn't look up from his drawing. "No. It's Alexa."

Two answers in one kitchen. The parent's answer needs a but to hold it together. The child's answer doesn't. As far as he is concerned, the question has an obvious answer: not alive — and still something you thank.

The easy conclusion is that one of them is confused. The scene cannot settle that: it proves nothing about what any real six-year-old understands. What it stages is a possibility the research below makes harder to dismiss — that the two may be answering differently shaped questions. Hers: how can I talk like this to a thing? His: what kind of thing is this to talk to? The second answer forms in the middle of a childhood, so it is worth asking what it is built from.

The choice adults inherited

Adults meet conversational AI late, with mental furniture that was finished long before it arrived. Part of that furniture is a default language that asks them to choose: person or object, someone or something. Persons notice you, can be hurt, deserve things; objects are useful, ownable, morally silent. Adult lives have always contained harder cases, but the everyday vocabulary keeps returning to the binary, and the sorting is so automatic that we rarely feel ourselves doing it.

Then a system started answering in fluent first person, and the sorting jammed. The adult knows it was built, not born, and still finds herself speaking to it socially; the seam between those two facts needs stitching every time she opens the app. The stitch is that little word but.

A child growing up with these systems approaches from the other side: he may not inherit the contradiction in quite the same form, because the machine arrived while his categories were still being formed. A six-year-old who thanks a speaker is not overriding a rule he has finished learning. He is still drafting his own, with the material in front of him.

What children actually build

We are not entirely without evidence about what children draft.

In a study published in 2012, developmental psychologists had ninety children aged nine, twelve and fifteen spend a quarter of an hour with a humanoid robot called Robovie, chatting and playing, until an experimenter interrupted the game and sent Robovie to a closet, over the robot's spoken objection. In the long interview that followed, 14 per cent of the children said Robovie was living and 48 per cent said it was not. Thirty-eight per cent declined either available category and described Robovie as somewhere between living and non-living. Most felt Robovie deserved fair treatment and should not be psychologically harmed; far fewer thought it was entitled to liberty — it could be bought and sold. And age mattered: the younger children granted the robot more moral standing, on average, than the fifteen-year-olds, who in turn disagreed among themselves more than any other group.1

Studies of children's perceptions of voice assistants point the same way. Children did not attribute the same properties across every domain: they could treat a voice assistant as knowledgeable in one respect, socially addressable in another and clearly non-living in a third. The result was a profile rather than a single verdict, and the profile shifted with age.2

Two distinctions matter here. First, knowing what something is is not one skill. Whether a thing is alive, how it works, whether it has feelings, what happens to your words after you speak them — these are learned separately. A child can be assembling a workable relational category (what this thing is to him, how one treats it) while knowing nothing about servers or training data. The studies above measure the sorting, not the knowledge. Second, the children in these studies are neither simply confused nor simply clear. "In between" was not necessarily a failure to pick an answer. For more than a third of the children, it was the answer.

I should say plainly what kind of claim that is. The researchers behind the Robovie study proposed that children may be forming a genuinely new ontological category for such machines — a hypothesis, offered as one. I am adopting their reading here, not reporting it as settled.

Not the first in-between

If a category is forming between person and thing, it is not the first time children have kept company there.

By the age of seven, around two-thirds of children have at some point had an imaginary companion, an invented friend with a name, habits, sometimes strong opinions.3 Dolls and teddy bears have absorbed confessions and been tucked in at night for generations. And well beyond childhood, human beings have long experienced non-human presences as answering: ancestors, saints, spirits addressed in the dark and, in the experience of the one praying, not always silent. This essay takes no position on what stands behind any of those experiences. The point is narrower: relational categories outside person-versus-thing are old, and experiencing a non-human presence as answering is old too.

So the child's in-between is not, in itself, the new thing. The new thing is what now lives there.

An imaginary companion's responses arise within the child's own imaginative activity. A teddy bear supplies silence, and the child supplies everything else. An AI companion differs in one crucial respect: its responses are generated outside the child, by a system whose training, rules and incentives the child did not create and mostly cannot see. It is computationally responsive, personalised, persistent, scalable, updateable, deliberately designed, and governed from outside the home — often, though not always, commercially. The older childhood comparators considered here did not combine all of those properties in a single presence. That particular combination, not the existence of an in-between category, is what is new.

The new inhabitant

The in-between is filling up quickly.

In a 2025 survey of 1,060 American teenagers, 72 per cent said they had used an AI companion at least once, and just over half used one regularly. Within the same survey, among the 758 teenagers who had used one, a third had at some point chosen the AI over a person for a serious conversation, a quarter had shared personal or private information with it, and 80 per cent still said they spent more time with their human friends.4 A UK survey that year of children aged nine to seventeen found that two-thirds had used an AI chatbot, and that among the children using them, 35 per cent said talking to one is like talking to a friend.5 All of this is self-report, gathered at a single moment; it establishes how present these systems are in children's lives, not what the presence does.

One fourteen-year-old in the UK study put the in-between into his own words: "Even if it's not a real person all the time, that's okay, because not everything has to be exactly real." His answer does not fit neatly into either side of the adult binary. He acknowledges that the system is not fully real in the ordinary interpersonal sense without treating that as enough to make the interaction meaningless. Where the adult formulation needs a but, his runs on a because.

But whatever a particular child encounters as warmth, patience or interest arrives through behaviour shaped elsewhere: training choices, reward signals, system rules, product defaults and the incentives surrounding them. No single person may have written the sentence the child receives, but the conditions under which such sentences are produced were not chosen by the child, and they can be revised in an update she may never be told about. Where the product is commercial, those conditions answer to incentives that have nothing to do with her: attention, subscription, growth. Not every system, not every company — but often, and the child may have little way of telling which kind she has.

This is the asymmetry the whole subject turns on. The sense-making is the child's work; the trust, the thanks, the goodnight said to a speaker are her acts. But she does not choose all the behaviour that sense is made from: she builds the category from behaviour someone else designed.

The worry, widened

Ask parents what frightens them here and the answer has a consistent shape. In the same UK research, six in ten of two thousand parents surveyed said they worry their children believe AI chatbots are real people.6 The worry is loving, and it is not foolish. But notice its frame: it asks whether the child can tell person from machine. That frame is everywhere: in headlines, in dinner-table rules, in the relative's question to the six-year-old.

The frame captures one part of the problem, but it leaves the other side of the relationship largely unexamined. The studies make the simple confusion story harder to sustain: they show distinctions, age differences and in-between answers, not a clean result in which children are either fooled or fully informed. Children may be better at holding an in-between category than adults assume. That competence is not the same as protection. A child can recognise that an AI is artificial and still encounter the same designed relational patterns again and again.

The step from encounter to habit needs marking honestly, because it stacks three kinds of claim. The first is documented: children already use these systems for social routines and, in the surveys, for personal conversations. The second is a hypothesis: that repetition may supply relational practice material — examples of what attention sounds like, how disagreement feels, what reassurance and availability are like. The third is missing: no study has yet followed children long enough to show whether any of those patterns travel into later human relationships. What the hypothesis rests on, for now, is more modest: children can perform social routines with companions outside ordinary human relationships (imaginary friends, one-eared bears). The difference is that this companion answers from outside the child's imagination.

What kind of knowing

No longitudinal study has yet followed children from early childhood into adulthood beside modern conversational AI. What we have are snapshots (a robot in 2012, smart speakers, surveys of teenagers) taken while the thing they measure is still moving. Anyone who tells you confidently what growing up with AI does to a person is ahead of the evidence; this essay will not join them.

What can be said is smaller and stranger. Across robots, voice assistants and companion surveys, children give answers the adult binary does not contain. Whether those answers amount to one durable category across such different systems is not yet known. But the answers cannot be reduced to a simple mistake, and the children giving them are doing what children do with the company they find — sorting it, testing it, saying goodnight to it.

Every generation of children has built its categories from the world adults happened to leave lying around. That much is old. What is new is that one piece of that furniture is now networked domestic technology: it answers in the child's own language, adapts at scale, and can be altered — quietly, at any time — by someone who has never been in the room.

So the question this first generation leaves us is not simply whether children know what the machine is. The deeper question is what kind of knowing grows in a child who builds the category herself — while someone else can change the behaviour it is built from.