When should children start using AI? Where we’ve landed, and why (Part 1 of 2)

  • Nick Alchin
  • September 28, 2026
Technology, Uncategorized
Back to blog

This post looks at our research, thinking and approach to AI. The next one is about integrity – how we guard against AI use in assessment, how well that works, and why we’re wary of an arms race.

(If you’d rather skip the research, go straight to section 2 or 3).

1 Some Research

Among parents, among staff, among students – some people are excited about the possibilities of AI, and some are very concerned. Often at the same time (I am one of those)! And I think we should all share the concern that students offload cognitive work when they should be building the capacity to do it themselves.

Let’s start with what is not new: offloading the effortful part of a task reduces what you learn from that task, and hence how you do on any subsequent assessment. That is the desirable-difficulties and retrieval-practice literature, which is about as robustly evidenced as anything we have in education, and it long predates AI (Bjork and Bjork, 2011 and I’ve written about it here, for example). What AI adds is not entirely new, but it is extraordinarily eadsy.

The idea that using AI measurably degrades thinking has something in it, though rather less than the apocalyptic coverage suggests. The initial widely-reported MIT study Your Brain on ChatGPT (Kosmyna et al, 2025) that started the ‘brainrot’ debate drew on a small laboratory experiment, and its authors have themselves pushed back on how it was covered: they have asked people to stop saying it shows brain damage or negative impact, and when asked directly whether their study shows AI makes us dumber, their answer is no.

That was the picture until this June, and the publication of a thirty-month study of 27,000 Chinese secondary students (Strömberg et al, 2026). Those who took up AI on their own saw their homework scores rise by 18%, and the time they spent on homework fall by 30%. That could just be efficiency – but monthly tests and high-stakes exams both fell by around 20%, which makes that implausible. Tellingly, the losses were among the 80% or so of AI users whose behaviour looked very much like outsourcing – homework finished suspiciously fast and scoring suspiciously well. It suggests that most of the students who took up AI on their own, with no one showing them how to use it well, drifted into outsourcing. Now, this is a working paper, not yet peer-reviewed, and it is only one county, where students chose for themselves whether to take AI up – so a limited study. But it is real data on real adolescents, and it is based on measurable outcomes, not a questionnaire. We should take it seriously.

So it doesn’t look good for students using AI alone, unsupervised, on their homework. Other studies are beginning to look more widely. A randomised trial (Bastani et al, 2025) gave one group unrestricted access and another an AI built to offer hints rather than answers. The first group did worse once access was removed; the second showed no measurable harm. So a plausible provisional conclusion is not that AI damages learning. It is that unstructured, unsupervised use does – for the desirable-difficulties reasons we’ve long known about. That is what we have to design around, and it is what the rest of this post is about.

2 Our Thinking

Some of you may be thinking: we heard all this about calculators. That’s fair, and broadly the calculator fear was overstated. The research through the 80s and 90s found calculator use didn’t damage problem-solving and often improved children’s attitude to mathematics. The schools that got the most out of calculators were the ones that redesigned the curriculum around them, deciding when and how they would be used. That was not accidental – we make sure we teach arithmetic before we hand over the calculator; the principle being to build a capacity before allowing a tool. 

Interestingly, economist Daniel Susskind argues that we should settle the AI question the way schools settled the calculator question: teach and examine every subject both with and without AI: “teach both, test both” as he puts it. IB maths already does this with calculators: some papers forbid them, others permit them. It’s too early to tell how well it works for AI, but we are giving it thought.

That said, our view is that AI and calculators are not equivalent. For schools, the issue is that AI is unusually frictionless, and cannot be fully policed in school or at home. A calculator, furthermore, only automates a narrow, well-specified sub-skill, and you can still see whether the child understands the problem. With AI you cannot so easily see if there is understanding at all. So we are not sanguine about the danger.

We do not think the solution is banning all access to AI, because a student who has never used AI under supervision has not been protected from it – more like left alone with it. The danger we identified is real, but so is the danger of students arriving at university having never developed the judgement to use AI tools properly. Banning it until the day they leave us simply guarantees they learn it somewhere with no supervision at all. It doesn’t solve the problem – it simply moves it from school to home; parents would be supervising it, without the tools or the curriculum or the colleagues to compare notes with.

3 Where We’ve Currently Landed

So what does that mean in practice? When and how should students be introduced to using AI? Reasonable people differ sharply here, and they differ inside our own community, and around the world: among parents, among our teaching staff. An LA district has banned AI for all K-12 students; New York City has paused student AI use up to Grade 8; the UAE does not allow it under 13; Singapore starts at Primary 4. Alpha School, a fast-growing private chain in the US, has gone furthest in the other direction with pre-K introduction.

  • No AI tools from K1 – G1. But that doesn’t mean kids don’t learn about AI – there are conversations about it, because they come up naturally when we talk to our students about how we find things out. So it means they find out about it before they use it.
  • In Grades 2 and 3, we have no standalone AI tools, but we have some educational subscriptions to digital tools that have some very narrow, limited AI tools built within them. These are designed for educational use and provide an opportunity to learn in a secure environment. The tools are always connected to a curriculum reason – misinformation and disinformation, for instance. AI per se is not an aim; the aim is to teach the dangers and limits of AI concretely rather than abstractly. A Grade 3 class creates a fake animal, and then asks: if you hadn’t created it, how would you know it’s fake? Very occasional use, a few times a year, not in assignments, never unsupervised.
  • In Grades 4 and 5 students begin structured interactions with AI through teacher-moderated platforms. There are no individual student accounts and no student data is shared. The teacher sets up a chatbot trained on that lesson’s resources, working within parameters they choose: for instance, don’t answer the student’s question directly, use questions to guide their understanding. A Grade 5 class recently used one to discuss the design tricks technology companies use to hold children’s attention, following a series of digital safety lessons. It gave students another way of showing their teacher what they understood, and the teacher could see and moderate every interaction. Used a few times a year, always inside a specific process.
  • In Grades 6, 7 and 8 the same teacher-moderated tools are used more often and across more subjects, as students develop the skills of interacting responsibly with AI. We also make limited use of Schools AI, and still provide no access to general AI tools.

We’ve looked at the evidence, at our own students and our own context, consulted the teachers who see this in classrooms every day and made a judgement. That is the basis for what follows. Of course, everything’s changing fast, so this is under constant iteration.

  • In Grades 9, 10, 11 and 12 High School students have access to Gemini through school accounts, which is a very different experience from the free open-ended version – it has more guardrails, and a guided learning mode that questions the student towards understanding rather than handing over answers. So the effortful work stays with the students, where it should be; in Design and Technology, for example, teachers have built AI characters that act as prospective clients, so student designers can meet them and develop a brief. Outside school they use many other tools, and we know it.

We are not out on a limb here. The IB has taken the same position, and published its reasoning: it will not ban these tools, because they will become more common and more powerful, because trying to ban or ignore them makes the eventual transition harder, and because education has to equip young people to deal with them competently and ethically. Like us, they are looking at assessment redesign on an ongoing basis.

Closer to home, the Singapore Ministry of Education has also landed in a very similar place. This was debated in Parliament in May, with members pressing the ministry hard on exactly the same issues – when should kids use AI, what safeguards, how do we stop over-reliance. The answer is a very similar gradient – their entry point is Primary 4, ours is Grade 4 – much the same age and in both cases under teacher supervision and with tools built for education. We are both also undertaking similar work on assessment. The risk MOE says it is guarding against is “cognitive offloading” – the same worry I opened with. 

This gradient of AI tool-use is defined in our K-12 Digital Literacy and Information curriculum, and also in a task framework that runs from no AI at all, through AI to plan, to collaborate, to edit, to create – with what the student does, what AI does, and how it must be cited, spelled out at each level. So we define when, and we define how. And the point is not that the AI needs teaching – children will pick it up anyway – probably better than many adults. But they will not pick up, simply by exposure, when not to use it, how to spot output that is confidently wrong, or the obligation to say you used it. That is judgement. It’s the hard, effortful part and it is learned from adults.

Starting young means we can build a framework of authenticity and integrity with our youngest. Then, when AI is introduced as a tool later on, it fits naturally into that framework. Alongside that we ask students to show the messy middle of a piece of work – on learning walls, in paper thinking journals, online – a habit that starts in the primary years and that we are extending upwards.

AI also forces us to redesign assessment so that it is AI-resistant, which is difficult and which we have not finished doing. But that is the world we’re in. It is and will be an ongoing process – and I’ll write further about what we have learned in due course.

One final point in this section. Let’s name the elephants in the room. AI systems consume significant energy and water; they carry the biases of what they were trained on; people were paid very little to label that training data, and much of the training material was taken without consent.

A dark side is not unique to AI; it is true of most of the systems we depend on. None of us are outside them. We fly to the conferences where we discuss emissions; we use phones built on questionable labour. But being implicated isn’t a reason to stop asking whether we should be using these systems at all – in fact, it’s most of why the question matters so much. We have made the judgement that the educational case for supervised use outweighs these costs, but they are costs nevertheless. The people who will eventually shape how these systems are built and governed are sitting in classrooms now. They will do that better if they understand the systems from the inside – the complexities as well as the harms – and are self-aware enough not to be captured by them.

Our protocol names these issues, and we ask students to weigh them: whether a particular use is worth its cost, and what they think should change.

Next Week: AI and Academic Integrity: How to Think About It and What To Do.

References

Share this article

Leave a Reply

Your email address will not be published. Required fields are marked *

Motivating Real Understanding in an AI Age

[This is more or less what I said at the ToddleAI Schools Showcase on 12 September, lightly tidied. It’s more a talk than a post, [...]

Learn more

Choosing Well and Losing Anyway: Loss Without Regret.

I was recently speaking with a student – call him Gill – choosing subjects for his IB Diploma. Most of it was easy, but the [...]

Learn more

What Individuals Can’t Do (Part 2 of 2)

Last time I said the world’s progress comes down to individuals. That’s true, but it’s nott the whole truth; here’s the other half. For as [...]

Learn more

Stay Updated!

Don’t miss out on the latest articles and insights. Subscribe now to get notifications straight to your inbox!

A reflection on educational and social issues, prompted by students, peers, community and world events.

Linkedin X-twitter
© 2026 Nick Alchin • All Rights Reserved

Made by Tokay Digital