How old should a child be to use AI? The labs disagree

Three of the biggest AI companies in the world sell broadly the same thing, and they cannot agree on how old a child has to be to use it.

Anthropic says 18. OpenAI says 13, with a parent’s permission until 18. Google will let a nine-year-old talk to Gemini, provided a parent flips a switch in Family Link first.

Same underlying technology. Three different answers. If you have ever read one of those age ratings and assumed somebody worked it out properly, read on.

What each lab actually says

Anthropic draws the highest line and does not hedge it. Its consumer terms read: “You must be at least 18 years old or the minimum age required to consent to use the Services in your location, whichever is higher.” When it launched Claude for Teachers on 14 July 2026, giving verified American K-12 teachers free access, it held that line. The offer follows Claude’s 18-and-over policy, it is not designed for direct student use, and it is available in the United States only. Anthropic’s route into your child’s schooling runs through the adult at the front of the room, which we wrote about in The AI labs want your child’s teacher.

OpenAI sets the floor at 13, and asks for a parent’s consent for anyone between 13 and 18, as the EU safer-internet service klicksafe summarised in March 2026. Since December 2025 the model itself has been written to treat that group differently. OpenAI’s Model Spec, the public document describing how ChatGPT is meant to behave, gained a set of under-18 principles: put teen safety ahead of other goals including intellectual freedom, steer teens towards real-world support, speak to them without condescension, be honest about what the model is and is not. Immersive romantic roleplay and first-person sexual or violent roleplay are ruled out for under-18s. An age-prediction model tries to catch minors who typed a false birthday at sign-up. Not everyone was convinced by the paperwork. One of the experts TechCrunch put it to, Steven Adler, said that “unless the company measures the actual behaviors, intentions are ultimately just words.”

Google goes youngest. Its own help page for families says a parent can give “your child under 13 (or the applicable age in your country)” access to Gemini through a supervised account, and that “a parent must enable access before their child under 13 … can use Gemini Apps with a supervised account.” A few things are withheld at that age: the Keep Activity setting, “Hey Google”, Voice Match. And to Google’s credit the same page is unusually plain about the limits. Its filters “try to limit inappropriate content from appearing, but they’re not perfect”. Gemini “can hallucinate and present inaccurate information as factual”. It “is not a person and cannot think for itself or feel emotions”.

The floor is a business decision, not a finding about children

Here is the detail that gives the game away. Google’s under-13 Gemini access is not available for supervised accounts in the European Economic Area, Switzerland and the United Kingdom. Nothing about a ten-year-old changes on the way across the Channel. The regulation does.

So the honest reading, and this is the desk’s view rather than a reported fact, is that these numbers track legal exposure and each company’s appetite for risk far more closely than they track anything published about child development. None of the three points to a study when it names its age.

A number arrived at that way can still be sensible. It just is not a rating, and it should not be read as one.

Singapore’s children are already below all three lines

The gap between the terms and the household is not small here. A study of 2,985 Singapore children aged 8 to 13, drawn from the Singapore Longitudinal Early Development Study and reported in May 2026, found more than half of eight-year-olds have used AI tools. By 10 it is more than seven in ten. By 13, more than nine in ten. The most popular tools were ChatGPT, Meta AI and Google Gemini.

Read that arithmetic slowly. The most popular tool among Singapore children aged 8 to 13 has a stated minimum age of 13.

Dr Xuejiao Chen, a senior scientist at A*STAR’s Institute for Human Development and Potential and one of the researchers, put the problem plainly: “I think a lot of parents do not understand how AI works, or how to safeguard their children when they use AI.”

Schools here draw their line somewhere else entirely, and for a stated reason. In a parliamentary reply on 6 May 2026, MOE said that for Primary 1 to 3, hands-on physical learning is prioritised and schools will not assign work requiring direct AI use. From Primary 4, pupils may use AI tools that are specifically designed for education, under teacher supervision, because research suggests that by around that age they have developed executive functioning skills such as planning and the ability to evaluate their own thinking. MOE also noted the obvious: “General purpose AI is also easily available online.”

That is an age line built on what a child can do. The app store’s line is built on something else.

What to do with a number that does not mean what it looks like

Not a ban. Three small shifts instead.

Treat the age floor as a legal statement and judge the tool yourself. Twenty minutes with it, asking the sort of thing your child asks, tells you more than the rating ever will.

Then ask which account your child is actually in, because the same brand ships very different products. A supervised Gemini inside Family Link is not the Gemini on your phone. A linked teen ChatGPT account is not an unlinked one, and OpenAI’s age prediction only works on the account the child really uses. If you have not set that up, our piece on what ChatGPT’s parental controls do walks through it.

And assume the gate is soft. As klicksafe noted, there is still no reliable age verification or consent mechanism behind any of this. A birthday field is a form, not a door.

(A disclosure, since it is our own patch: Mentus AI builds AI mentors for children, so we have a commercial interest in how parents think about age and AI.)

The labs have each told you, quite precisely, what they are willing to be responsible for. The distance between that number and what your child is already doing is not a gap in the terms. It is the part that was always yours.