Before eight, the honest answer is supervised use only, and mostly not a chatbot. The AI a young child meets is the assistant that answers the moment they speak: in a smart speaker, on a tablet, through the TV. Use it with them rather than handing it over, keep it off an account of their own, and treat every filter as leaky, because the companies say so themselves.
That last part is the bit worth holding on to. Every control below is worth setting. Not one of them is a reason to stop paying attention.
Where do young children actually meet AI?
Not where most parents expect. The question I get asked is about ChatGPT, and the thing already running in the house is a voice assistant.
Ofcom's Children and Parents: Media Use and Attitudes report 2026, published on May 21, 2026, from 5,008 interviews with UK parents, found that among parents of 3 to 7 year olds whose child goes online to do creative activities, just over half, 52 percent, say their child has used AI during those activities. Read that carefully, because the qualifier does real work: it is not 52 percent of all young children, it is 52 percent of the ones already drawing, making and building online. Within that group, AI is not an upcoming decision. It has happened.
And the route in is rarely an app a child chose. It is the speaker on the kitchen counter that now answers in paragraphs instead of reading out a fact. Amazon's Alexa+, its generative version, is free with Prime and otherwise $19.99 a month, available to all customers in the United States and Canada, and in early access in the UK, Brazil, Mexico, Italy, Spain, Germany, Austria and France. Nobody installs that. It arrives.
So the useful question before eight is not which app to allow. It is what happens to a five year old when the thing in the corner answers instantly, every time, and never once asks them to have a go first.
Is a smart speaker safe for a young child?
Safe enough to have in the house, with a child profile switched on, and not safe to treat as a babysitter or a tutor. Two specific things are worth knowing.
The first is what it does with your child's voice. In May 2023 the FTC and the Department of Justice charged Amazon with breaking children's privacy law over Alexa, and Amazon agreed to a $25 million civil penalty. The regulators' allegation was that Amazon "prevented parents from exercising their deletion rights under the COPPA Rule, kept sensitive voice and geolocation data for years, and used it for its own purposes," including to improve the Alexa algorithm. The settlement required Amazon to delete inactive child profiles, delete transcripts of children's statements, and stop training algorithms on that data.
That case is three years old and the order changed Amazon's obligations, so it is history rather than a live warning. I include it because it answers the question parents actually have, which is whether a recording of their four year old is sitting on a server somewhere, and because it establishes the principle better than any vendor page will: deletion rights only matter if somebody uses them. Go and look at what is stored under your child's profile, once.
The second is what the kids mode does. On Amazon's own account of it, devices in Amazon Kids mode block voice shopping and mature content such as explicit lyrics by default, and the Parent Dashboard lets a parent set time limits, approve which contacts a child can call, and restrict categories of content while leaving reading unlimited. Amazon describes the experience as spanning preschoolers to pre-teens, with read-alongs, math games and science trivia.
All of which is genuinely useful, and none of which touches the thing I would actually watch for. A content filter decides what a child is allowed to hear. It has no view on whether they are being asked to think.
Can a child under 13 use Gemini or ChatGPT?
Sometimes, and it depends on the company and on where you live. The detail differs enough between products that we went through what each company requires and what those numbers buy you separately, so here is only the part that matters before eight.
Google is the one that reaches young children, because it runs through the account a parent already set up. Google's guidance on Gemini Apps for supervised accounts states that "you can give your child under 13 (or the applicable age in your country) access to Gemini Apps," and that "a parent must enable access before their child under 13 can use Gemini Apps with a supervised account." So this is a decision a parent makes, not a default to accept.
Google is also unusually straight about the limits, and I would rather quote it than soften it. Its own page says: "Our filters try to limit inappropriate content from appearing, but they're not perfect. Your child may encounter content you don't want them to see." It warns that Gemini "can hallucinate and present inaccurate information as factual." And it says something I wish more companies said out loud: "Gemini is conversational and can feel like a human. It's important for you to help your child understand that it's an AI tool, not a person to confide in."
One geography note that saves UK and European parents a hunt through menus: Google states that Gemini Apps "are not available for supervised accounts in the European Economic Area, Switzerland, and the United Kingdom." If you are reading this in Britain, that particular decision is not yours to make yet.
What is appropriate at each stage before 8?
The band is short and it changes fast, so it is worth splitting into three.
Under 3: none of it, and this one is not a close call. The American Academy of Pediatrics, writing for parents on healthychildren.org on June 3, 2026 about its current digital media policy, puts it plainly: infants under 18 months learn best from real-world interactions, and heavy solo screen use can affect developing language and social skills. A conversational assistant is the most convincing fake conversation ever built. A child who is still learning that talking is something you do with a person does not need one. The device change: turn the wake word off in the rooms where they sleep and play, or simply do not create a profile.
3 to 5: together, on your device, in short bursts. This is the age for the good version of AI, which is a shared thing you do for four minutes and then stop. Ask it for a story about your child's own toy. Ask it what sound a pangolin makes. Then turn it off, because the value was the two of you being delighted, not the answer. Say out loud that it is a machine, every time, in the same flat way you would say the oven is hot. The device change: set up a child profile or kids mode, so that shopping is off and the assistant is not answering everything a small voice throws at it.
6 to 7: still supervised, and now you can start the habit that matters. At this age a child can hold the idea that a confident answer might be wrong, which is the single most valuable thing they can learn about this technology and the hardest to teach later. So check one answer together, out loud, on purpose. Let them catch it being wrong. The device change: on Android, open familylink.google.com or the Family Link app, tap your child, then Controls, then Gemini, then Gemini Apps, and make that choice deliberately rather than inheriting it.
A young child does not need a better answer. They need a longer moment before the answer arrives.
At 8 the question changes shape, because friends, accounts and school devices arrive at roughly the same time. That is a different article, and we have written it: the guide to the years before a phone picks up exactly where this one stops. If what you are really weighing is whether your particular child is ready rather than what the bands say, the three signs I would look for, which I call the Ready Test, apply at any age and are a better guide than a number.
What I See in Schools
The pattern primary teachers describe to me most often is not a safeguarding incident. It is a five or six year old who has stopped guessing.
Guessing is how young children learn. A child who says "I think the moon is made of cheese" and gets to be gloriously wrong in front of a kind adult is doing the exact cognitive work that later becomes reasoning. The teachers I train say that the children who have had a lot of instant-answer time at home are quicker to say "I don't know" and wait, and slower to offer the wrong answer that would have taught them something. I want to be careful here, because that is what experienced teachers report to me rather than a measured finding, and I have not seen a study that isolates it. But it is consistent enough, across enough schools, that I would not wait for the study before acting on it at home.
It is also the clearest case I know for the principle I keep coming back to: let AI help with the doing, and keep the thinking with your child. Before eight, almost all of the thinking should still be theirs, which means almost none of the doing should be the machine's.
What to say to your child
Short, plain and repeated beats one big conversation. These are the lines I would actually use.
At 3 to 5, when it answers something:
You could say: "That's the computer talking. It doesn't know things like you and I know things. It just guesses really fast."
At 6 to 7, when it gets something right:
You could say: "It sounded sure, didn't it? Let's find out if it was right. How could we check?"
At 6 to 7, if they ask it something that should have come to you:
You could say: "You can ask me that one. I might get it wrong too, but I'll think about it with you."
And once, in a calm moment, the line that matters more than any setting:
You could say: "If anything on there ever makes you feel funny or scared, come and get me. You'll never be in trouble for telling me."
What none of this can guarantee
The filters are imperfect by the manufacturers' own admission, quoted above. A child profile does not stop a young child using an adult's unlocked phone, which is how most of them meet these tools in the first place. A time limit says nothing about what happened inside the time. And none of these controls exist on a friend's tablet, a grandparent's speaker, or the device in the back of a car.
Two things take precedence over anything in this article. If your child's school has its own rule about AI on school devices or school accounts, that rule wins on their equipment, whatever you have set at home. And if you are worried about your child's speech, language or development, that is a conversation for your pediatrician, your family doctor or the school's special education lead, not something to resolve by adjusting a setting. AI has nothing to offer that question.
What to ask the school
For an elementary school child, or a primary school child if you are in the UK, three questions get honest answers and take about a minute at the door.
- Is any AI switched on in the tools my child uses here, and on whose account? A school Google or Microsoft account is administered by the school, so the school decides, not you.
- What have staff been trained to do with it in the early years specifically? Early years is where general AI guidance fits worst, and the answer tells you how thought-through the policy is.
- Who do I talk to if I want it off for my child? A school that can name the person has a real position. A school that cannot is still working it out, which is worth knowing kindly rather than crossly.
If the answers are vague, that is usually not indifference. It is that nobody has given the staff anything to work with yet. Training school staff on exactly this is the work I do with schools, and it is a fair thing to forward to a school leader who is being honest with you about not knowing.
For tonight, though, the smallest useful thing is a walk around your own house. Name every device that answers when spoken to. For each one, decide: on with me, on with a child profile, or off. That is the whole decision, and it takes about ten minutes.
I write a Sunday newsletter that keeps teachers and parents up to date on this without the panic. Parents are very welcome.
Dan Fitzpatrick is The AI Educator: a former teacher, assistant headteacher and school trustee who now helps schools and governments work out what AI means for children. More about Dan.
