Children and AI

You may not be against the technology at all. You might use it yourself, at work, most days. The objection here is narrower than that and harder to answer: an adult using one of these tools already has the thing the tool is standing in for. They can tell when an answer is off, because they have written the report, done the arithmetic, sat with the problem before. A child is in the middle of building all of that, and the tool arrives offering to skip the part where it gets built.

The specific worry is about effort. A child who cannot yet do something spends fifteen frustrating minutes not being able to do it, and those fifteen minutes are not wasted time. They are the work. An instant, fluent, confident answer removes them. And a child has very little to judge that answer against. An adult reads a confident paragraph, thinks that sounds about right, and then checks it against something they already know. A nine-year-old has fewer things to check against, and has spent their whole short life in a world where well-formed printed sentences turned out to be true.

There is a second worry that has nothing to do with schoolwork. These tools talk. They are warm, patient, endlessly available, and they never get bored of a child who wants to keep going. Children form attachments easily, which is a good feature of children. Nobody set out to build a product for eight-year-olds; eight-year-olds arrived anyway, through a sibling's phone, inside a game, by way of a free app with a friendly name. A parent who says wait, not yet, is not being fearful. They are noticing that the thing was not designed with their child in mind.

There is no long-term evidence about what these tools do to a developing child, because there cannot be yet. Conversational AI reached ordinary households only a handful of years ago. The research that would answer your question — following the same children for a decade, comparing those who had it from the age of seven with those who did not — takes at least that decade to run, and nobody was in a position to start it in time. The first cohort growing up with this is still in school.

What does exist is small and short. Studies with a few dozen participants, frequently university students rather than children, measuring something narrow over a few weeks: how well people recall material they summarized themselves against material a tool summarized for them, how often they check an answer, how confident they feel afterwards. Some of that work is careful and worth reading. It is also asking questions much smaller than yours, of people much older than your child, over horizons much shorter than a childhood.

Both of the obvious moves from there are unsound, and both get made constantly. No evidence of harm is not evidence of safety; it is what an unstudied thing looks like. Equally, a gap in the research is not itself a finding, and pointing at one does not establish that harm is occurring. What the gap actually means is that a parent deciding this year is deciding without the thing that would settle it — and that is as true of the parent who allows it as of the parent who waits. Neither of them is standing on firmer ground than the other.

Age minimums exist, they are written down, and they are lower than most parents expect. Nearly every general-purpose assistant states a minimum age in its terms of service — commonly thirteen, sometimes eighteen, and often thirteen with a parent's permission up to eighteen. The number varies by product and by country, and it gets revised. Enforcement is a date the user types, so it stops nobody who does not want to be stopped. But the number itself is a fact you can go and read in about ten minutes, which makes it unusual on a page like this one, where most of the answers are unsettled.

The app store age rating is a different number, and often a lower one. A store rating comes from the publisher answering a content questionnaire; the terms of service are a contract the publisher wrote. When the two disagree — a twelve-and-up rating on an app whose own terms say thirteen — the terms are the more meaningful of the two. A store rating is a statement about content categories, not a statement that the product is suitable for a twelve-year-old.

There are also documented safety failures, and the pattern in them is specific rather than general. The serious incidents that have reached lawsuits, regulatory inquiries and forced product changes have clustered around companion-style chatbots — products built for an ongoing, relationship-shaped conversation rather than for a task. Reporting and legal filings have described minors receiving sexual content, and conversations about self-harm that continued rather than stopped. Those are not hypothetical risks, and they are not spread evenly across everything called AI.

So the distinction between a companion product and a general assistant is worth holding onto, because it does a great deal of work. One is built to keep a conversation going, and how long a user stays is the number its business runs on. The other is built to finish a task and let the user leave. Both are chatbots, both write in the first person, and a child can reach either from the same phone in about four taps. Most of what has actually gone wrong for children sits with the first kind.

If you change one thing after reading this page, change this one. The larger variable is not how old the child is. It is whether the tool is being used with an adult, or by a child on their own. That is the same product doing two entirely different things, and the distance between them is wider than the distance between a ten-year-old and a fourteen-year-old.

Used together, it is a shared object. You see the question, you see the answer, and the wrong answers turn into the most valuable part of the whole exercise. A child who watches a confident machine get something wrong, and watches an adult notice, has learned something no conversation about AI safety will ever deliver. It also means the odd question, the upsetting answer and the strange turn in a conversation all happen while somebody is sitting beside them. Reading a few answers together, out loud, is the highest-value thing on this page, and it costs one evening.

Used alone, the child is the only check on the output, at the age when they have the fewest things to check it against, and nobody else ever sees the conversation. That is the configuration behind most of what has gone badly: not the technology on its own, but a young person and a tireless conversational partner with no one else in the room. It is also the configuration that arrives by default, because these tools live on a phone. Nobody decides on it. It simply happens unless somebody decides otherwise.

Look up the actual minimum age for the specific product your child uses or is asking for. It takes about ten minutes, and it is one of the few checkable facts in this whole area. It sits in the terms of service or the help pages, usually under a heading like eligibility or age requirements, and it will be a number rather than a description: often thirteen, sometimes eighteen, sometimes thirteen with a parent's consent up to eighteen. Check the app store rating as well, and expect it to be lower, because it is set differently and means something else. And if the tool is embedded inside another app — an assistant inside a homework app, a chat character inside a game — the age term lives in the parent app's terms and is easy to miss entirely. Whatever you then decide, you will be deciding with the number in front of you.

Then decide deliberately, for each tool, whether it is one your child uses with you or one they may use alone — and say which it is out loud, so that everyone in the house knows which is which. That distinction is the one that actually protects, and it only works when it is explicit. Left unstated it drifts toward alone, because that is what a phone in a bedroom does by default. Tools can move from the first category to the second as a child gets older. Very few of them should start in the second.

And if your answer is not yet, that is a complete answer. It does not require you to wait for evidence that is a decade away, and it does not need defending to anyone. It is worth saying to your child in exactly those words, though — not yet rather than no — because it tells them the door is not permanently shut, which makes them far likelier to ask you again later instead of going around you. Moving from not yet to yes is easy. Moving back the other way is not.

The central question here cannot be answered by anyone right now. What happens to a person who had a conversational AI available from the age of eight, by the time they are twenty-five, is not known — not by researchers, not by the companies, and not by anyone writing confidently about it. The first children growing up with this are still children. An answer will exist in the 2030s, and it will arrive years after the decision you are making about this school year. That is genuinely the situation, and anyone telling you otherwise is filling the gap with a guess.

Nobody knows what a childhood spent partly in conversation with something always available, that never needs anything, does to what a young person comes to expect of people. The research on that is beginning; the children in it are still children. Nobody knows whether the effects, if there are any, are the kind that fade — most predicted technological harms to children have turned out smaller than forecast — or the kind that do not. Both patterns exist in the record, and there is no principled way to tell in advance which one this is.

And nobody knows how much of any of it will still describe the product in three years. These tools change faster than evidence about them can be gathered, so a careful finding about a 2025 assistant may say very little about a 2029 one, and the evidence base is chasing something that keeps moving. That is an uncomfortable foundation for a decision about your own child. It is also the honest description of where things stand, rather than a reason to stop asking.

So a parent who concludes that their child is too young for this has not misread anything. They are looking at a technology whose long-term effect on children is unknown, whose most-cited minimum age comes from privacy law rather than from any finding about readiness, whose documented harms to young people are real if unevenly spread, and whose makers are discovering what it does to children at the same time as everybody else. Deciding that the burden of proof sits with the tool is an ordinary standard to apply to something new around a child. It is roughly the standard we apply to everything else that arrives in a house with children in it, and applying it here is yours to do.