By the time we sit down together on 17 July, you’ll already have had this conversation in some form, more than once. About whether to ban AI. About whether to integrate it. About whether to mark essays differently. About what to tell parents. About how to keep students safe. About what your competitors are doing. About whether you’re moving too fast or too slow.
I want to suggest in this short note that none of those is the central question.
The central question, the one your governance is actually responsible for, is harder. It’s what AI is doing to thinking.
A school is in the business of building minds. The instruments of that work are subtle, and they’ve been refined over centuries. The essay. The problem set. The close reading. The lab. The rehearsal. The practice. None of these is really designed to produce the work you see at the end. Each is designed to do something to the student while they’re doing the work. That something has, by the time the student leaves your schools, become part of them in ways neither you nor they can fully name.
That’s what’s under discussion when we discuss AI in schools. The finished work is now cheap. A student can produce a competent essay in ninety seconds without doing the thing the essay was for. The work looks the same as what the student would have produced. What the work was a vehicle for may not have happened at all.
That’s the question I’d like to walk you through on the morning. Not whether we should use these tools, or how to catch cheaters, or what to tell parents. Something deeper. What is AI doing, right now, to the formation of the minds your schools exist to form, and what do we do about it?
The evidence is more interesting than the headlines
Two views dominate the public conversation. The boosters say AI will revolutionise learning, multiply teaching capacity, and democratise expertise. The doomers say AI is rotting children’s brains and that schools should ban it. Neither is what the evidence actually shows.
It depends. Whether AI helps or harms a learner depends on how it’s used, on what the learner does while using it, and on which cognitive capacities were being built in the first place. The same model, in the same task, can produce a large positive effect, a large negative effect, or no effect, depending on whether the student does the work or has the work done for them.
Three recent studies make this concrete. In a 2024 randomised trial, fifty experienced physicians were given access to GPT-4 alongside their usual reference tools, or to the references alone. The two groups performed at almost exactly the same level on a set of clinical cases. The model on its own, with no doctor, performed substantially better than either group. The doctors had access to the better answer and couldn’t reliably tell when to take it.
In a 2025 field experiment with about a thousand high school maths students in Turkey, students were randomised to use a chatbot during practice or not. The chatbot group did forty-eight percent better while they had access. Then access was removed and they were tested. The chatbot group did seventeen percent worse than the control group. Practice with the tool hadn’t built the skill.
In a 2025 study from Microsoft Research and Carnegie Mellon, researchers surveyed three hundred and nineteen knowledge workers across more than nine hundred real AI-assisted tasks. They found that workers’ cognitive role had shifted. People were doing less producing and more verifying, less solving and more integrating, less authoring and more signing off. Not necessarily a bad shift. But the kind of shift that’s hard to notice from inside it, and one that depends on the verifier having the discrimination to know when an output is wrong.
These aren’t the only findings, and the literature has grown large enough that the convergent picture can be put in one line. Once anyone can generate competent work for almost nothing, competence stops being the thing worth having. What becomes scarce, and valuable, is judgement.
Why the public conversation has missed it
The public conversation has mostly been about three things. Cheating. Productivity. Future skills. Each one is real. None of them is the central question.
Cheating is a problem of detection and policy. It’s being addressed seriously, including by the new mandatory academic integrity course being run for the 2026 Queensland Certificate of Education cohort. But it’s also, in a sense, the wrong framing of the deeper risk. The deeper risk isn’t that students who used AI will be caught and punished. It’s that students who used AI won’t be caught, because there’s nothing to catch, and they’ll arrive at adulthood having handed in fluent work for years without the formation the work was supposed to produce.
Productivity is also real. A modern school does a lot of work that can be done faster with these tools. Lesson plans, vendor proposals, draft policies, parent communications. There are real efficiencies on offer. But none is the central question for governance, because governance isn’t, principally, about efficiency. It’s about what the institution is for.
Future skills is the framing I’m most cautious of. The argument runs that students need to be familiar with AI because their working lives will be saturated with it. The argument isn’t wrong, but it’s too cheap. Familiarity with a tool isn’t a substitute for the capacities the tool is now eroding. Telling a fifteen-year-old to be fluent with ChatGPT, while AI does for them the writing and reading the next forty years of their professional life will depend on, is a trade most parents wouldn’t knowingly accept if it were spelled out for them.
The shift in framing
Schools have always built judgement. The instruments of school work, again, aren’t arbitrary. The slow drafting, the comparison of arguments, the gradual feel for what counts as evidence and what counts as a guess, the lived experience of being wrong and noticing it and adjusting. These are how judgement is formed. Not from the work, but from the doing.
The deep risk of AI in schools isn’t that students will cheat. Students have always found ways to cheat. The deep risk is that students will do the work and hand in the work and not develop the judgement, because the work has been done by something else, with everything looking right on the outside and the inside missing. And the gap won’t show up in any current assessment, because the work looks the same either way. It’ll show up much later. At thirty, when these students are running organisations, or sitting on boards, or making medical decisions, or voting.
That’s what I’ll be asking the room to think with me about. I’ll offer a frame, drawn from the cognitive sciences, that may be useful at the level of governance. There’ll be specifics. There’ll be evidence. There’ll be things to protect, things to build, and things to measure, in ways that translate across the classroom and the board table.
You don’t need to do anything in advance to be ready. But if you have a few minutes between now and then, I’d suggest one observation. Pick any task that AI can now produce competently for you. A draft email. A summary of a paper. A meeting agenda. A plain-language version of a complex document. Watch yourself use the tool, just once, with full attention. Notice the moment the tool produced something. Then watch what you did with it. Did you send on the version it gave you, or did you change it? Did you read it carefully, or did you accept it because nothing on its surface invited you to push back?
That moment, multiplied many times a day across your classrooms and at your board table, is the territory we’ll walk together. The fact that you can already feel something about it, even if you can’t quite name it, is the reason the conversation is worth your time.
Jason