How much AI is too much for a child? The honest answer, and the better question

Here is the least mystical way to describe what that adult is noticing.

Finding a word is not looking something up in a fixed store. It is an act of retrieval, and retrieval is trainable in the same unglamorous way as anything else — the more often you successfully reach for something, the faster and more reliably you reach for it next time. This is one of the better-established findings in memory research: the act of pulling information out strengthens the path to it, considerably more than reading it again would.

Which means the reverse is also true. Every time something else supplies the word, you skip a rep. Not a catastrophic one. Just one. The spelling correction you never saw, the sentence you did not have to construct, the point you did not have to formulate before you could ask for it to be formulated.

That is a much plainer description than 'AI is rotting your brain', and it is also much more likely to be right. Nothing is being damaged. Something is simply being practiced less, and practice is what made it quick.

It is worth being honest that this specific chain — AI use, less retrieval practice, harder word-finding — has not been demonstrated end to end in a study. The most-quoted research linking AI use to weaker critical thinking is a survey of adults who reported on themselves; it found a strong correlation and it cannot show which way the causation runs. People who already think less may simply use AI more. The mechanism is solid. The proof that it is what is happening to any particular person is not.

The word-finding part deserves a caveat that is genuinely reassuring, and it is the kind of thing nobody tells bilingual adults.

Bilingual speakers have more tip-of-the-tongue moments than monolinguals. This is a well-replicated finding, and the explanation is almost boringly mechanical: it is called the frequency-lag hypothesis. If you split your speaking across two languages, each individual word in each language gets used less often than the equivalent word does for someone running everything through one language. Word frequency strongly affects how easily a word comes to hand. So each one sits slightly further away.

Bilinguals also name pictures more slowly in laboratory tasks, and the gap is widest for uncommon words — exactly the words that go missing mid-sentence.

So if you are bilingual and you feel yourself groping for a word, some meaningful portion of that is the ordinary cost of speaking two languages, and it was there before any of this. That is not a reason to dismiss the feeling. It is a reason to suspect that AI use may be sharpening something that already existed rather than creating it — and that the two stack, because both work through the same lever: how often you actually produce the word yourself.

It is also worth ruling out the dull explanations before the interesting one. Poor sleep, stress, illness and simply getting older all produce exactly this symptom, and all of them are more common than anyone likes to admit.

Now put that adult next to a seven-year-old, because the difference between them is the whole argument.

When an adult offloads a task to AI, they are handing over a skill they already built. The path exists; it is getting less traffic. Give it a month of writing your own emails and it comes back.

A child may not have built it yet. The effortful, irritating, slow work of finding the word, planning the sentence, holding the argument in mind long enough to write it down — that work is not an obstacle in front of the learning. In young children it substantially is the learning. One Brookings researcher puts it about as plainly as it can be put: cognitive development, like learning to ride a bike, cannot be outsourced.

That is a serious argument from serious people. It is also, at the moment, a hypothesis. A paper published this summer, criticizing the habit of transferring adult findings onto young people, was blunt about it: nothing yet links a particular mode of AI use to long-term development in children. We would rather tell you that a worry is plausible and unproven than sell it to you as settled.

In 2024 a group of researchers ran a trial with around a thousand students in years equivalent to US grades nine to eleven, in maths classes at a high school in Turkey. Three groups. One studied the normal way with notes and a textbook. One got a plain ChatGPT interface. One got the same underlying model, but set up to give hints and ask questions rather than hand over answers, and primed with the mistakes students usually make on that topic.

While the AI was available, both AI groups did dramatically better than the control group on practice problems. That is the part everyone expects.

Then the AI was taken away and they sat an ordinary exam. The students who had been given plain ChatGPT scored meaningfully worse than the students who had never had AI at all. The students who had used the hint-giving version scored no differently from the control group — no harm.

Same model. Same students. Same amount of time. The only thing that differed was whether the tool handed over the answer or made them reach for it. When the researchers looked at what students had actually typed, the plain-ChatGPT group had mostly just asked for solutions. The other group had asked for guidance and then attempted the problem.

That is the clearest result in this whole field, and it is not about minutes. It is about whether the child did the thinking or watched the thinking happen.

There is a Massachusetts Institute of Technology study on essay writing that has been reported everywhere, usually badly. It is worth knowing what it actually says, because it is the closest thing to a measurement of the feeling described at the top of this article.

Researchers had people write essays either with ChatGPT, with a search engine, or with nothing at all, while wearing EEG caps. The ChatGPT group showed weaker and less distributed patterns of brain activity during the task. They also reported feeling less ownership of what they had written — and in the first session, most of them could not correctly quote a line from the essay they had just produced, while almost everyone in the other two groups could.

Three honest caveats. It is a preprint, not peer-reviewed work. The authors have gone unusually far out of their way to ask people not to describe it as brain scans, as damage, or as evidence that AI makes anyone less intelligent — their own words. And every participant was an adult university student, in one narrow geographic pocket, writing one kind of essay in twenty-minute sittings. There were no children in it, and no dose was tested.

What it is reasonably good evidence of is the thing the adult at the top of this article noticed: that work you route through a machine does not stick to you the way work you did yourself does.

It would be easy, and dishonest, to stop at the worrying half. Some of the better-designed studies with young children found benefits.

In one, around a hundred and twenty children aged three to six had a story read to them either with back-and-forth dialogue or straight through, by either an adult or a conversational AI. The dialogue mattered enormously. The partner mattered much less — comprehension with the AI was comparable to comprehension with a person. A related project built an AI character into a children's science programme so kids could answer it out loud, and the children who could talk back learned more of the science than the ones who just watched.

But the same researcher found the other half of it too: children talk less to a machine than to a person, put less effort in, take less initiative, and are quicker to give up and accommodate a misunderstanding rather than push to be understood. In one study with four to eight year olds, children had more communication breakdowns with an AI partner and made fewer attempts to repair them. They just adjusted themselves to the machine.

Her own framing is that AI is different from, and complementary to, human interaction. An extra source. Not a substitute for the person on the sofa.

The American Academy of Pediatrics quietly did something useful in January: it moved away from hour-based screen limits altogether, toward a set of questions about the child, the content, and what the screen is displacing. That reframing is exactly right here, and it survives the fact that we do not have a number.

So, four questions that are answerable at your kitchen table tonight:

If you want something to watch for that is more informative than elapsed time, watch for dependence. In that same US survey, among children who used AI for schoolwork, a noticeable minority agreed they had difficulty starting or finishing homework when they could not use it — and that figure was substantially higher among the children using it every day.

That is the signal. Not the clock. A child who can do the work without it and chooses to use it is in a different position from a child who has stopped being able to start.

The other thing worth knowing, because it tends to surprise parents: a meaningful share of nine to twelve year olds in that survey had used AI to talk about feelings or personal problems. Pediatric guidance is consistent on this one — the line to hold is that it is a tool, not a companion, and younger children in particular are prone to treating a thing that talks back as a friend.

We publish books and browser games, so we have an obvious interest in you choosing a picture book over a chatbot, and you should weigh what we say accordingly.

What we would say regardless is that the most protective thing in all of this literature is not a rule. It is an adult in the room. Dialogue is what made the AI reading study work, supervision is what the pediatricians keep returning to, and a hint rather than an answer is what made the difference in the one large trial that measured it. All three are versions of the same thing: somebody is still doing the thinking with them.