Thirteen, if you mean the age your suppliers have already agreed to. OpenAI states that ChatGPT is not meant for children under 13 and requires parental consent from 13 to 18. Google scoped the Gemini app in Workspace for Education to teen users aged 13 and over, then opened Gemini in Classroom to students of all ages from August 10, 2026. That is the contractual answer. It is the only one of the three age limits your school is already operating under that arrives with a number, and on its own it will not survive a parent's second question.
The other two carry no number at all. The task sets one: whatever the work is building in a child, if it is still under construction, a machine doing it takes the learning with it. Your own school sets the other, and it is the one nobody writes down, the age at which an adult here could tell what a child actually did. Where the three disagree, the most restrictive governs until somebody writes down why it should not.
That is a position you can say out loud to a parent, a governor or an inspector. A single number is not, which is why the schools that publish one spend the year defending it.
What age should students be allowed to use AI?
The answer is 13 where the terms govern, task by task where the learning governs, and lesson by lesson where your own supervision governs.
Most policies get the first part right and then mistake it for the whole job. Thirteen tells you who has agreed to what. It tells you nothing about whether a fourteen-year-old should be drafting a history essay with a chatbot, and nothing about whether you would know if she had. The other two lines resist being written as rules at all. They can only be written as judgments somebody is accountable for, which is why they drop out of the document.
A leadership team that can say which of the three it is invoking, on which decision, has an age policy. One that publishes a number has a press release.
The Three Age Lines
The Three Age Lines are the three age limits every school is already operating under, whether it has noticed them or not:
- The contractual line. Drawn by the supplier, in terms your school accepted without choosing. Ask: what age do the terms say, and has anyone here read them this year? Move it by choosing a product built and licensed for that age, or by putting an adult at the keyboard.
- The developmental line. Drawn by the task, not the birthday. Below it, the learning a task builds is still under construction, so handing the task to a machine takes the learning with it. Ask: what is this task building, and is it still under construction? Move it by changing the task so the machine is not doing the part that teaches.
- The supervisory line. Drawn by your school's own capacity to see. It is the age at which an adult here could tell what a child actually did with the tool. Ask: could an adult here tell, and how soon? Move it by changing when and where the work happens.
A policy that draws one line and assumes it covers the other two is not an age policy. It is one line doing three jobs badly. The three lines exist whether or not you draw them, and they rarely fall in the same place. I offer the frame as a way of thinking rather than a tested instrument.
The question is live this term because of New York. On September 2, 2026, the city told nearly 600,000 children that they would not be using generative AI at school this year. The moratorium runs from 2-K to the end of eighth grade, covering two thirds of the system's students. Companion chatbots are prohibited across all grades. Teachers keep their access. Since then, the question leaders have put to me is not whether New York was right. It is where the line should fall in their own school.
The contractual line: the age you have already agreed to
The contractual line is the minimum age written into the supplier's terms, and in most schools it is quietly being crossed by tools that staff recommend in good faith.
OpenAI's position is unambiguous. Its help documentation states that "ChatGPT is not meant for children under 13, and we require that children ages 13 to 18 obtain parental consent before using ChatGPT." It goes further for schools: where ChatGPT is used in education with children under 13, "the actual interaction with ChatGPT must be conducted by an adult." Google drew its line in the same place, and has since moved it. When it opened the Gemini app to students through Workspace for Education, the announcement of November 8, 2024 scoped access to "teen users (ages 13+ or the applicable age in your country)." Then on August 4, 2026 Google announced that from August 10 it would open Gemini in Classroom to "K-12 and higher education students of all ages who have already been granted access," lifting the 18-and-over restriction that had applied there. Access is on by default, and an administrator who wants under-18s excluded has to put them in a separate organizational unit and switch it off.
That pair is the contractual line behaving the way contractual lines behave. Two products from one company, two different ages, and the one that moved moved by default, in a release note, over a summer. A school that wrote "13 and over" into its policy in 2024 and has not looked since is describing a setting it may no longer have.
OpenAI has since added a second kind of contractual line, one the supplier draws on your behalf. On August 18, 2026 it described ChatGPT for Teens, an experience that applies when its system estimates a user is under 18 or the user states an age between 13 and 17. Safeguards are on by default, parents can link accounts and set Quiet Hours, and Study Hours can be set with Study Mode on by default. OpenAI describes the product as meant for learning outside the classroom, and points schools toward a separate offering, ChatGPT for Teachers, which gives educators and districts institution-managed access. My reading is that a teen mode protects the child using the tool at home and approves nothing in your classroom. It also means the age a supplier applies to your students may be estimated rather than declared, so the age a child typed and the age the tool treats them as can differ, and your school chose neither.
Something awkward follows. A primary or elementary school debating whether to let nine-year-olds use a general-purpose assistant is debating a question its suppliers have already answered. The choice actually available is narrower: an adult works the tool while children watch and direct, or the school buys a product built for that age group, or it does neither and accepts that it is operating outside the terms it agreed to.
None of this makes thirteen a safety threshold. It is a legal and commercial boundary, not a developmental finding, and treating it as evidence about children is the second mistake in this area. What it does tell you is who has agreed to what, which is precisely the question a parent will ask first.
It is also a line children cross without us. The Common Sense Media Census, published June 8, 2026 from a survey of 1,204 children aged 9 to 17, found that 86% use AI and nearly a quarter use it daily, while more than 40% said no parent or guardian had ever discussed AI safety with them. A school that sets its line at thirteen and stops there has settled its paperwork and left most of its nine to twelve year olds using these tools with no adult in the conversation at all.
The developmental line: what is still being built
The developmental line is not about age. It is about whether the thing a task builds is finished.
This is where the Handover Question does its work. The Handover Question is the one test I ask school leaders to put to every task children are set: if a machine did this for them, what would they lose? If the answer is nothing, hand it over and use the time better. If the answer is the learning itself, protect it, and be able to say why.
Notice what that test does not contain: a birthday. A sixteen-year-old structuring an extended argument for the first time is at the same point in that skill as a nine-year-old composing a paragraph for the first time. The developmental line moves by subject, by task and by pupil, which is why it cannot be set once at the front of a policy and left there.
England's Department for Education has built the same idea into what it expects of products rather than of children. Its generative AI product safety standards, published January 19, 2026, ask developers to "only show a full solution after a genuine learner attempt," and to monitor and report to teachers "the rate of requests for cognitive offloading and the amount of cognitive offloading delivered." Those standards are written for suppliers, not schools, and the department says only that schools may find them helpful. They show where the argument has moved: the serious question is no longer whether a child is old enough, it is whether the product is doing the part of the work that was supposed to teach them something.
The supervisory line: whether an adult could tell
The supervisory line is the age at which your school can actually see what a child did with AI, and it is almost always the line nobody has drawn.
This one is not about the child. It is about you. A task set in a supervised lesson, on a school device, with the process visible, sits above the supervisory line at almost any age. The same task set as homework, on a family tablet, sits below it at every age, because nobody in the school can tell what happened. Schools routinely write policies that permit at fourteen what they cannot observe at any age, and then treat the resulting confusion as a student integrity problem.
The regulation now being written assumes somebody is doing the seeing. The same DfE standards note that services must use "highly effective age assurance" to keep children in the UK from encountering the most harmful content, under the Online Safety Act. Age assurance at the product end only works if someone at the school end knows which children are using which tools and can check. That is a job with a name attached, which is the same argument I make about what a school should require before approving an AI tool at all.
What New York decided, and what it did not
New York drew a grade line, which is the easiest kind of line to enforce and the hardest kind to defend on its own.
The detail is more interesting than the headline. The moratorium covers student-facing generative AI from 2-K through eighth grade for one year. Companion chatbots go across all grades. For a maximum of 50,000 high school students in general education classes, five named pilots continue: Quill, Edia, Brisk Teaching, Playlab and Intel AI-Ready Schools. They run in a maximum of five classes per high school, under the direct supervision of a trained educator, with weekly limits such as no more than 15 minutes a week on Quill and no more than two assignments per marking period on Playlab. High schoolers also get twice-yearly AI literacy classes. Teachers keep AI for planning and operational work. A Technology in Schools Coalition will assess the moratorium and publish a report with recommendations for future school years.
Read those pilot terms as the supervisory line written down. A cap on classes per school, a trained adult in the room and minutes counted are what it looks like when a system declines to let an age do the work that supervision has to do. That reading is mine, not the city's.
The exceptions are the part worth studying. The city says exceptions will be granted for assistive technology used by students with disabilities, and press coverage of the announcement adds multilingual learners and students in career readiness programs such as computer science. Every one of those is a developmental judgment wearing the clothes of a grade rule. A student using assistive technology to reach the curriculum is not offloading the learning, they are getting to it. A grade line alone was not sufficient even for the system that drew it, and the exceptions are where the real policy lives.
If your leadership team is weighing a restriction of its own rather than a line inside one, the prior question is how to make a restriction decision you can defend, and it should be settled before the age question, not after.
What I Have Learned From Working With Schools
The age line is almost never written down, and leaders usually discover their own only when a parent asks them what it is.
Speaking at education conferences internationally, the question I am asked most often once the session ends is no longer which tool a school should choose. It is at what age. The same question arrives in the same words from systems with completely different regulations, which is itself a finding: nobody has been given a defensible answer, so everybody is improvising one.
Across the schools I work with, three patterns repeat. The line is inherited rather than set, usually from whichever platform the school already licenses. The contractual line is discovered late, often in the middle of a complaint, which is the worst possible moment to read the terms for the first time. And the supervisory line is never discussed at all, because it is the only one of the three that is about the adults.
Three things I say in every session on this. Attach the age in your policy to a task rather than to a birthday, or you will spend the year defending an arbitrary number. If you cannot say who set your current line, you have not set it, you have accepted someone else's. And a line you cannot supervise is not a line, it is an intention, and students work out the difference faster than staff do.
When the three lines disagree
When the three lines disagree, the most restrictive one governs, and moving it requires changing something real rather than rewriting the policy.
There are only three honest ways to move a line. Change the product, to one built and licensed for that age. Change the task, so the machine is not doing the part that teaches. Change the supervision, so the work happens where an adult can see it. Anything else is not a decision, it is a permission slip written to the school by itself.
This gives a governing board no single number to publish, and a number they cannot defend is worse than a rule they can. It also asks more of staff than a grade cut-off does, because task-level judgments belong to the person setting the task. The way through that workload is a short list of task types agreed once for a department rather than a judgment made afresh every lesson.
What to do this term: the five-step checklist
Start by finding out which line you are actually operating under, because almost every school has one and almost none has chosen it. Five steps, in this order:
- Read the terms. Take the three or four AI-capable tools your students touch most and write the minimum age from each supplier's terms in one column beside the year groups that use them. Where the column contradicts itself, you have found your contractual gap. This takes an hour and settles arguments that have run for a term.
- Check your defaults. Open the admin console for each of those tools and confirm which age groups have access today, not which age groups your 2024 policy says they have. For consumer tools that staff and families use, also ask whether the supplier now applies a teen experience based on an age it estimates.
- Put the Handover Question to two tasks per subject. Not twenty. Mark each one hand over, protect, or redesign.
- Test one homework task the same way. Homework is where the supervisory line fails first, so it tells you most for the least work.
- Write one sentence for parents saying what your school allows at what age and why, and check that a parent could act on it at the kitchen table without calling the school. If they could not, the policy is about you rather than about them, which is the point I make about talking to parents about AI.
Keep teaching about AI at every age while you restrict its use at some of them. A child who has never used it and never understood it is the worst of both decisions, which is the distinction underneath what AI literacy actually means in a school. Where this decision falls among everything else a new term brings is the ground the first 90 days of leading AI change covers, and the seven dimensions of AI readiness show whether your school's supervision is strong enough to hold the lines you draw.
If your leadership team is setting this line for the year ahead, this is the kind of decision I help schools and trusts settle through AI strategy sessions for schools: not which number to publish, but which three lines you are holding and who owns each one. A line only holds if it lives somewhere people read, which is the argument underneath an AI strategy people actually use. If you want this thinking as it develops, it goes out first in my newsletter for school and system leaders.
Sources and further reading
- Office of the Mayor, City of New York, "Mayor Mamdani and Chancellor Samuels Put Students First with Nation's Broadest Generative AI Moratorium in Schools," September 2, 2026. nyc.gov
- OpenAI, "Introducing ChatGPT for Teens: Built for learning, backed by protections," August 18, 2026. openai.com
- OpenAI, "Is ChatGPT safe for all ages?," OpenAI Help Center, read October 15, 2026. help.openai.com
- Google Workspace Updates, "Expanding access to the Gemini app for teen students in education," November 8, 2024. workspaceupdates.googleblog.com
- Google Workspace Updates, "Gemini in Google Classroom is expanding to users of all ages, with contextualized Gemini starter prompts for students," August 4, 2026. workspaceupdates.googleblog.com
- Department for Education, "Generative AI: product safety standards," GOV.UK, published January 19, 2026. gov.uk
- Common Sense Media, "Common Sense Media Releases Inaugural Annual Study on AI Use by Tweens and Teens," June 8, 2026. commonsensemedia.org
Dan Fitzpatrick is the founder of The AI Educator and speaks internationally on what AI means for schools and the people who lead them. More about Dan.
