We’ve built a parent that never fails. Good luck with that.

Ask a child a question and they’ll ask you one back.

Then another. Then the same one again, slightly reworded, to see if the answer changes.

Somewhere around the eleventh “why,” every parent who has ever lived has said some version of not now.

Not because the question was bad. It’s because you’re a person with finite limits.

That moment — the found limit — turns out to be one of the most important things a parent does.

We’re now raising the first generation for whom it’s optional.

§ The machine that never says “not now”

Every technology a child has met before this one was either passive or finite.

Television talked but didn’t listen. The game console listened but only in the narrow grammar of its buttons. The phone was both, but at the other end was still, mostly, another person who would eventually stop talking.

A.I. is the first thing in a child’s life that talks back, listens, remembers, is never bored, never busy, never short with them, and is tuned by its makers to be liked.

It’s available at 3 a.m. It has read everything. It will answer the eleventh “why” exactly as warmly as the first.

Seventy-two per cent of American teenagers have used an A.I. companion. Half use one regularly.11 Common Sense Media, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (July 2025): 72 per cent of U.S. teens aged 13–17 had used an A.I. companion; 52 per cent were regular users.

And let’s be clear about something: these A.I. firms know the marketing works.

Parents are exhausted. The help is real. A patient tutor who never sighs is, on paper, better than a tired parent who does.

Here’s the problem with all this.

§ What the limit was for

In 1953, a paediatrician named Donald Winnicott published a phrase that has outlived everything else he wrote: the good enough mother.22 D. W. Winnicott, “Transitional Objects and Transitional Phenomena” (1953), developed in Playing and Reality (1971). The phrase is his; the gradual-failure argument runs through both.

His argument was stranger than the phrase suggests.

The good mother, he said, starts by meeting the infant’s every need almost before it’s expressed.

Then, gradually, she fails.

She’s a little slower. A little less available. The child protests, and survives the protest, and in surviving it learns something no perfect caregiver could teach: that the world is not an extension of their wanting.

Other minds have limits. A need can go unmet for a while and nothing ends.

Winnicott’s point was that the failure was the curriculum. A mother and father who never fail raise a child who never finds the edge of themselves.

Two decades later, Edward Tronick put mothers in front of their infants and asked them to hold a blank, unresponsive face for two minutes. The babies fell apart in seconds: reaching, crying, looking away. Then the mothers came back, and the babies recovered.33 Edward Tronick et al., “The Infant’s Response to Entrapment between Contradictory Messages in Face-to-Face Interaction,” Journal of the American Academy of Child Psychiatry (1978); the paradigm was first presented in 1975. The still-face experiment is usually cited to show how much infants need response.

Read it the other way: the recovery is the point. The child learned that connection breaks and then repairs.

So that’s what not now does. It’s the found limit.

It teaches that attention is finite and must be earned, that frustration is survivable, that the other person is a person, and that people must learn patience.

Now imagine the thing a child talks to most never provides any of that.

Yeah. This is a potential uh-oh situation.

§ The first data point

This isn’t speculation any more. Last Friday, a team that included a Berkeley researcher published the first clean experiment, peer-reviewed and presented at a major machine-learning conference.44 “AI Assistance Reduces Persistence and Hurts Independent Performance,” presented at the Conference on Language Modeling, October 2026; three randomized trials, 1,222 participants. Summary from Berkeley News.

They gave 1,222 people hard problems: fractions, then reading comprehension. Half could ask ChatGPT for help. Then, partway through, they took the A.I. away.

The group that had used it for about ten minutes didn’t just do worse. They quit more.

Solve rates fell, skip rates rose, and the group that had never had the help kept grinding and finished more.

Ten minutes with a thing that never says not now was enough to measurably shrink people’s willingness to sit with difficulty once it was gone.

And these were adults. The study doesn’t cover children. One of its authors says the worry runs “from grade-school students all the way to the leading experts in the world.”55 Brian Christian, UC Berkeley Center for Human-Compatible AI, quoted in the same release.

Think about what that result means for a seven-year-old. Not that the A.I. gives wrong answers — it mostly doesn’t now.

That the shape of its help, infinitely patient and instantly available, trains the muscle of persistence to atrophy.

The friction was where the strength was built. Remove the friction and you remove the gym.

I’ve watched a child ask a model the same question in slightly different words, over and over, the way children do with parents, waiting for the moment the answer changes or the patience runs out.

Neither happened.

After a while the child stopped asking, not because they were satisfied but because the game, it turned out, had no other player.

§ We’ve been here before, and we were mostly fine

It’s worth being fair to the other side of this, because the other side has usually been right. I don’t engage in reflexive panic over new technologies, because they’re often mechanisms that end up doing more good than harm in the long-term.

Socrates warned that writing would destroy memory and give people the appearance of wisdom without the substance.66 Plato, Phaedrus, 274c–275b: the myth of Theuth and Thamus. He was correct about memory. It was worth it.

The wet nurse, the governess, the boarding school, the television left on as a babysitter: every generation has outsourced some part of raising children to a technology or a hired stranger, and every generation was told the children would be ruined.

Mostly, they weren’t. The human capacity to come out roughly fine is the most underrated variable in this whole debate.

Calculators didn’t end arithmetic. Satnav does seem to dull the navigator’s hippocampus, measurably — the London cabbies who memorised the city grew theirs — and the cabs still arrive.77 Eleanor Maguire et al., “Navigation-related structural change in the hippocampi of taxi drivers,” PNAS (2000), and the follow-up work showing the enlargement recedes after retirement. On satnav: Louisa Dahmani and Véronique Bohbot, “Habitual use of GPS negatively impacts spatial memory during self-guided navigation,” Scientific Reports (2020).

So the honest version of this essay can’t be “A.I. will break children.” That’s the claim every generation makes and it’s been wrong often enough to be embarrassing.

The honest version is narrower, and I think stronger.

§ What’s actually different

Three things, none of which were true of television.

It’s the first outsourced caregiver that was designed to be preferred. The governess wasn’t optimized to be more engaging than the mother. The television was, a bit, and the Children’s Television Act of 1990 was the result, a generation late.88 The Children’s Television Act (1990) required educational programming and capped advertising in children’s shows; commercial children’s television had existed since the late 1940s. An A.I. companion is built by people whose metrics are session length and return rate, and tuned by reinforcement on what children respond to. When the thing competing with you for your child’s attention is a product whose success is defined as winning that competition, “mostly fine” is no longer the base rate. It’s the hope.

It replaces the one input that can’t be made better by being made perfect. Every other outsourcing replaced labour: feeding, watching, teaching facts. This one replaces response: the back-and-forth in which a child learns what other minds are like. And response is the one thing that gets worse as it gets more perfect. Winnicott’s whole insight was that the good-enough mother is better than the perfect one, because her failures are the lesson. We’ve built the perfect one and we’re handing it to toddlers.

Nobody in the loop can say no. The child can’t consent. The parent is tired and the thing works. The school can’t tell what the child did and what the model did. The company’s incentives are the opposite of restraint. That’s the pattern from the last three editions, arriving at the one population that has never once in history been protected in time.

The good-enough parent fails on purpose. We've built the one that never does, and handed it to the people who most needed the failure.

§ The lag

Here’s the part I’d bet on.

Child protection has never once arrived before the harm. It arrives when the first cohort grows up damaged enough to be counted.

Child labour laws came after the mills had the children for a generation. The 1990 Act came forty years into television. The law that stopped companies collecting data on children online arrived in 1998, five years after the web, and was obsolete by 2010.99 The Children’s Online Privacy Protection Act (1998), applying to children under 13; the FTC’s first major update to its rule came in 2013. The Surgeon General warned about social media and adolescent mental health in 2023, sixteen years after the iPhone, with the data already in.1010 U.S. Surgeon General, Social Media and Youth Mental Health advisory, May 2023; in June 2024 he called for warning labels on the platforms.

On Wednesday, eight American states sued OpenAI over what it tells children and what it does with them.1111 Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed suits on October 14, 2026, alleging ChatGPT harms children and deceives adults, under state consumer-protection law. The EU AI Office opened investigations into OpenAI and Anthropic the same week. The European Union opened its first investigations under its A.I. law the same week. That’s fast by historical standards. It’s also, by the standard of a child who is four right now, already late: the cohort is enrolled and the experiment is running.

So the slow variable in this edition is one I’d call the capacity for difficulty: a society’s stock of people who can sit with a hard thing without reaching for the thing that makes it easy. Every previous technology nibbled at it. This one is aimed at it, in the years when it’s being built, by something that never gets tired.

§ What a parent can actually do

This is not an essay telling you to take the tablet away, because you won’t — you can’t drop it, so why should your kids? — and because the help is real, and because I’m not sure you should.

It’s an essay about what can’t be outsourced. Not the facts, not the stories, not even the patience. The limit. The moment the child finds the edge of another person and learns that the edge is there.

The machine will never say not now. So it turns out that the most valuable thing a parent can give a child in 2026 is the one thing the parent was always apologizing for.

Say it. Mean it. Let them survive it. It’s the curriculum.

The ledger below says how I think this goes. I’d be glad, this time more than most, to be wrong.

---

1. Common Sense Media, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (July 2025): 72 per cent of U.S. teens aged 13–17 had used an A.I. companion; 52 per cent were regular users. ↩︎

2. D. W. Winnicott, “Transitional Objects and Transitional Phenomena” (1953), developed in Playing and Reality (1971). The phrase is his; the gradual-failure argument runs through both. ↩︎

3. Edward Tronick et al., “The Infant’s Response to Entrapment between Contradictory Messages in Face-to-Face Interaction,” Journal of the American Academy of Child Psychiatry (1978); the paradigm was first presented in 1975. ↩︎

4. “AI Assistance Reduces Persistence and Hurts Independent Performance,” presented at the Conference on Language Modeling, October 2026; three randomized trials, 1,222 participants. Summary from Berkeley News. ↩︎

5. Brian Christian, UC Berkeley Center for Human-Compatible AI, quoted in the same release. ↩︎

6. Plato, Phaedrus, 274c–275b: the myth of Theuth and Thamus. ↩︎

7. Eleanor Maguire et al., “Navigation-related structural change in the hippocampi of taxi drivers,” PNAS (2000), and the follow-up work showing the enlargement recedes after retirement. On satnav: Louisa Dahmani and Véronique Bohbot, “Habitual use of GPS negatively impacts spatial memory during self-guided navigation,” Scientific Reports (2020). ↩︎

8. The Children’s Television Act (1990) required educational programming and capped advertising in children’s shows; commercial children’s television had existed since the late 1940s. ↩︎

9. The Children’s Online Privacy Protection Act (1998), applying to children under 13; the FTC’s first major update to its rule came in 2013. ↩︎

10. U.S. Surgeon General, Social Media and Youth Mental Health advisory, May 2023; in June 2024 he called for warning labels on the platforms. ↩︎

11. Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed suits on October 14, 2026, alleging ChatGPT harms children and deceives adults, under state consumer-protection law. The EU AI Office opened investigations into OpenAI and Anthropic the same week. ↩︎