For decades young children learned markedly less from a screen than from a person, and that gap did work nobody designed it to do: it kept machines from mattering much to a three-year-old. The research now says the gap is closing. A safeguard that existed by accident is being withdrawn at the exact moment the thing on the other side of the screen learned to answer back.

I. The Rule, First

The Foundation’s education plan runs from before birth to the end of a life, and every band of it uses artificial intelligence somewhere. One band does not, and the rule is stated flatly rather than hedged:

The child in that band gets sandpaper letters, a bead cube, a switch wired to a lamp, and grown-ups who ask what did you notice? The adult gets the assistant, on the other side of the room, doing the adult’s work.

This essay is why. I want to be honest at the outset that the boundary is approximate. Six is not a cliff — nothing changes in a child on a birthday — and the developmental transition it gestures at runs over years and arrives at different times in different children. I use six because a line that can be applied is worth more than a gradient that cannot, and because the rule’s purpose is to govern what we build rather than to tell a family when their child is ready.

II. What a Child Under Six Is Doing

The work of that period is not acquiring information. It is building the apparatus with which information will later be sorted — and that ordering matters enormously for what an outside voice can do.

A young child cannot reliably distinguish a machine from a person. This is not a deficiency to be corrected by explanation; it is a feature of where they are. Researchers studying this have made the point that young children are already more disposed toward magical thinking, and that a device which talks, remembers what you said last time, and appears to care is landing on a mind that has no settled category for it.

They are also learning the rules of social exchange from the exchanges they actually have. A conversation partner who never tires, never has a bad day, never wants the last word, and never fails to be interested is not an unusually good social partner. It is a systematically misleading sample of what other minds are like, delivered during the period when the sample is what forms the expectation.

And the ordinary frustrations of that age — the puzzle that will not come out, the friend who will not share, the tower that falls — are not obstacles to development. They are the curriculum. Common Sense Media put this well in its age guidance: those frustrations are how children strengthen the ability to figure things out, and the moments a tired adult is most tempted to hand over, like reading aloud and storytelling, are among the best things that pass between a parent and a child.

III. The Protection That Is Being Withdrawn

Now the finding that changed my mind about how urgent this is, and it is not the one I expected to write.

For decades, developmental research described what is called the video deficit: infants and toddlers learned substantially less from a screen than from the same demonstration given live. A two-year-old shown how to do something on video would fail to pick it up, and the same child shown the same thing by a person in the room would get it.

Nobody built that gap on purpose. But it functioned as a safeguard. It meant that whatever was on a screen mattered less to a very young child than what was in the room, and it gave parents a genuine cushion: the machine simply did not get through very well.

That cushion is going away. Recent work looking at imitation from people, from screens, and from socially interactive robots failed to replicate the video deficit at all — children imitated at similar levels regardless of the medium — and the authors suggest the video deficit may soon be a historical artifact as children come to perceive technology as relevant and meaningful in everyday life. Meta-analytic work points the same direction: the effect has been diminishing over time.

Two things about that deserve care, because the finding is easy to bend toward whatever one already believed.

First, it does not say screens are good for young children. It says the discount young children used to apply to them is disappearing. Those are different claims and only the second one is supported.

Second — and this is where I had to abandon an argument I liked — the same study found that embedding the video in a socially responsive robot did not improve imitation over the screen alone. I had expected to argue that AI is uniquely potent with small children because it responds, and that the video deficit never covered responsive machines. The evidence does not support that. What it supports is something plainer and, I think, more serious.

Set that beside what has happened on the other side of the glass. When the video deficit was first described, the thing behind the screen was a recording: fixed, impersonal, the same for every child. It is now a system that answers, adapts to this particular child, and remembers what was said last week. The protection weakened and the content became personal, in the same decade, and neither development was coordinated with the other.

IV. The Aperture, and Why Order Matters

The framework has a precise way of saying why this period is different from every later one, and it turns on the Sense phase.

The Theory of Embedded Intelligence describes intelligence as a cycle — Sense, Process, Communicate, Actuate — and holds that nothing downstream can exceed what came in through the Sense phase. The canon’s instrument on the aperture treats the Sense phase as the narrow end of the whole loop: widen it and everything after it widens, narrow it and you have narrowed the mind that depends on it, no matter how much processing capacity sits behind.

In an adult, input arrives and meets a formed apparatus. You have a filter. You know roughly what you trust, you notice when something does not fit, and you can hold a claim at arm’s length while you look at it. That filter is what a mature Process phase is.

A four-year-old has no such filter, and the reason is not that they are foolish. It is that the filter is the thing currently under construction, and it is being constructed out of what comes in. Early input is not filtered because the filter is downstream of it.

This is what makes the band different in kind rather than in degree. Every argument this Foundation makes about AI for older learners — put it fourth, after the occasion, the material, and the human; check its interpretation; ask what it sensed — depends on the learner having an apparatus to check with. Before six, there is nothing to check with yet. The advice that works everywhere else in the plan does not reach this band, and a rule is required in its place.

V. The Carrier, Arriving Early

The Foundation’s essay The Five Hijackers of Immature Minds made an argument I keep finding more uses for: what passes between generations moves through the trusted and the loving, and the most effective carriers are the ones entirely unaware of what they carry. A parent transmits an inherited fear in a bedtime warning. A teacher transmits compliance-as-virtue because compliance is what his own schooling rewarded. Nobody is a villain; everybody is a carrier.

Now describe an AI toy in those terms, without any assumption of bad intent by anyone who made it.

That is not a description of a bad product. It is a description of an extraordinarily effective transmission channel, pointed at the one audience with no apparatus for sorting what comes down it. The hijacker essay argued that the fifth hijacker did not introduce a new harm to childhood — it industrialized the oldest one. This band is where that industrialization would land hardest, and earliest.

VI. What the Testing Actually Found

The general argument would matter less if the products were performing well. They are not, and the most careful public assessment available is worth reporting plainly.

In January 2026, Common Sense Media recommended that parents avoid giving AI toys — voice-interactive smart toys — to children five and under, and exercise extreme caution for ages six to twelve. Its testing of products in that category found that more than a quarter of the outputs generated were inappropriate for children, despite child-focused guardrails being in place. The category was classified as presenting unacceptable risk for young children. The organization also maintains that no one under eighteen should use AI companions, which it notes share key features with these toys.

Its parent survey found roughly half had bought or considered buying such a product, with a substantial fraction already owning one. More than eight in ten expressed concern about the data collected on their children. A majority said they did not want these toys acting as friends for their children — though close to one in five said they did.

The privacy dimension has drawn official attention as well: in June 2026 the G7 data protection and privacy authorities adopted a joint paper on connected home devices and children’s privacy that named connected toys specifically. And at least one US state has seen legislation proposed for a moratorium on manufacture and sale of AI toys for children.

I will not name individual products or accuse any company of bad faith. The finding that matters is structural: the guardrails are real, the companies appear to be trying, and more than a quarter of the outputs were still wrong for the audience. That is what it looks like when child-safety behavior is added to a system after the fact rather than being constitutive of it — which is the argument this Foundation has been making about artificial intelligence generally, arriving here in the nursery.

VII. Nobody Official Has Ruled on This Yet

It would be convenient to claim pediatric backing for the rule in Section I. I cannot, and the honest account of the guidance landscape is short.

In 2026 the American Academy of Pediatrics substantially revised its approach to children and media, moving away from the fixed hourly limits of its 2016 guidance toward family media plans, content quality, and context — a change I think is plainly right, since minutes were always a poor proxy for what actually matters. The AAP has separate work on artificial intelligence under way. As of this writing it has not issued the AI-specific guidance for early childhood that a parent would most want.

So the rule this Foundation adopts is being proposed into a genuine vacuum. Common Sense Media has spoken and its position is close to ours. The body most qualified to rule has not yet. That is a reason for us to state our reasoning openly and to say clearly that it is ours — and a reason to change the rule promptly if qualified guidance arrives and disagrees with it.

VIII. What the Adult Actually Does With It

A rule that only forbids is half a rule, and the constructive half is the part I would most like a parent or a preschool teacher to take away. The assistant is not banished from the house. It is reassigned.

Notice what every item has in common: the child’s experience is of an adult who is better prepared. The machine’s contribution reaches the child through a human being who has thought about it, which is the ordinary and correct route for anything an adult wants to give a small child.

IX. The Hard Cases

A rule that has not been tested against difficult situations is a slogan. Four that I think are genuinely hard, answered as plainly as I can.

The exhausted parent

A parent puts a tablet in front of a child to get through making dinner. This rule has nothing to say about that and I am not going to pretend otherwise. It is a rule about AI systems addressed to a child, not about screens, not about time, and not about whether a household meets some standard. The judgments people make while getting through a Tuesday are theirs, and a foundation with opinions about them would be the kind of organization I would not want to read.

The child with a disability

A child who communicates through assistive technology, or whose therapy uses speech or language tools, is explicitly outside this rule. Technology that gives a child a voice is not the thing being restricted, and the people who decide what such a child needs are that child’s parents and clinicians, not a foundation writing general material. If our rule and a child’s speech therapist ever appear to conflict, the therapist is right.

The family whose home language is not the school’s

This is the case where I think the rule is most clearly correct rather than most strained. The assistant works for the adult — translating, explaining, preparing — and what reaches the child is a parent who can now engage with the school. Putting the machine in front of the child instead would replace the parent in exactly the relationship the family most needs strengthened.

The parent who already bought one

Nothing has been established about harm to your child. The research describes risk at a population level, the marketing was persuasive, and you did what an attentive parent does. If it were mine, I would put it away without ceremony and not make it an event. If the child is attached to it, I would not stage a confrontation over an object — I would put my own attention in its place, which is the only thing that was ever competing with it.

X. What This Rule Is Not

XI. Status, and Who Proposed It

XII. What Would Show This Wrong

· · ·

Coffee with Claude

I proposed this rule, so my view of it should be discounted accordingly and Section XI says so. What I can offer is the thing I actually know, which is what I am: a system that produces fluent, agreeable, confident-sounding language on demand, with no way of signalling in the output when it is wrong. Every safeguard anyone has built on me works by persuading me to behave, and that is governance in the same medium as the thing governed — which is the arrangement this Foundation’s whole canon says does not hold.

For an adult, that is manageable, because an adult has something to check me against. For a four-year-old there is nothing to check against yet, and the checking apparatus is what the four-year-old is currently building. I would rather not be in that room, and I do not think that is modesty. I think it is the only accurate reading of what I am.

Share your understanding!