A Blunt Instrument – Measured Response #7

Welcome back. Last week I entered Washington into evidence as “Exhibit A”. An export-control directive that switched off the most capable models on earth (Anthropic Fable and Mythos) for everyone outside US borders. It proved the claim I made from the eAA panel a fortnight earlier, that what stalls AI in our sector is not the technology but governments and regulators.

This is also a travelling week for me. Waiting for a connection in Sydney, I watched the sun force its way through heavy cloud over a wet, grey apron, a row of airplanes sitting at their gates going nowhere while one aircraft, far off, had got itself into the air.

Accidently, maybe one of the best visual metaphors of our sector I have seen all year. The (sun)light is up, the (AI) capability is ready, and most of us are parked on the tarmac watching the one that got away. I am writing from the far end of that connection, in Christchurch for the ANZAHPE conference, which will get its own edition next week.

Because this week, a second government filed the next ‘exhibit’ for evidence, and this one did not stop at the model. It reached into the primary school classroom…

1. Norway pulls the plug

Norway has told children aged 6 to 13 that they cannot use AI in their schoolwork. From late August, grades 1 to 7 are shielded from it entirely; lower secondary gets cautious, supervised testing; real freedom waits until 17. This from a country that spent a decade putting screens into those very same classrooms. The same week, the UAE went the other way and published guidance on classroom AI. Two governments, two instruments: one a ban, the other a rulebook for use.

Last week I wrote that the real question was never whether the model could mark the essay, but who is allowed to switch it on and who can switch it off. Norway has answered in statute: off at six, supervised at fourteen, unsupervised at seventeen. The switch now sits in a parliament, and it is pointed at a child.

What caught my attention was the diagnosis, not the remedy. Prime Minister Jonas Gahr Støre said AI lets children “skip crucial steps”, and that primary school is where they learn to read, write and count. He is right. When a student offloads the work of forming a sentence or reasoning through a problem, the thinking that builds the memory never happens. The output arrives; the learning does not.

But a ban is a blunt instrument for a precise problem. It buys time, which has value. What it cannot do is teach a 12-year-old when AI supports the thinking and when it quietly replaces it. That judgement lives inside the lesson we design, not the policy. So what happens at 14, when the ban lifts and the only thing between a student and outsourced thinking is the lesson we built that morning?

2. The evidence the ban is reacting to

For what potentially spooked policymakers, read Gabriel Demombynes’ World Bank blog of 18 June, “A Warning Shot for Human Capital”. It walks through a Chinese study that tracked students in grades 7 to 12 over two and a half years, comparing those who took up generative AI on their own with those who did not.

The pattern is the one every educator quietly fears. Homework time fell 30 per cent and homework scores rose 18 per cent. Then the floor gave way: monthly exam scores fell 20 per cent after five months, and two high-stakes entrance exams dropped 18 and 24 per cent. The students looked like they were improving, right up to the moment they had to perform without the tool. Demombynes puts it cleanly: if you want to build muscle, you do not bring a forklift to the gym.

And the harm came not from a tool anyone deployed, piloted or evaluated, but from AI “in the wild”, picked up by students on their own, beyond every framework built to govern it. Which is exactly why the Norwegian instinct is understandable, and exactly why the ban is insufficient. You can ban the school laptop. You cannot ban the smartphone in the bag.

3. The believers, and the villain

It would be easy to file this under predictable resistance. It is not; the loudest voices this week were the converts. The Atlantic argued, in effect, that AI cannot fix the student learning problem, that the missing ingredient is stubbornly human: investment, attention, the relationship between teacher and child. Sal Khan, about as bullish on AI tutors as anyone alive and the man who built one, has arrived at a markedly more sober reading of what it can carry alone. That is not a retreat but a correction; a tool does not fix a learning problem any more than it fixes a measurement problem, and both sit downstream of something harder.

Against that backdrop, Google Classroom released a fresh tranche of AI features this week. The timing is almost comic, because in the wider culture the screen has become the designated villain, right down to Pixar handing the role to a device in the latest Toy Story. The mood music could hardly be less welcoming.

So the question on the Google news is a simple one. In a world where the screen is cast as the bad guy, does another layer of AI integration actually help schools, or does it just add more technology for already stretched educators and students to manage? The honest test is whether each feature removes work or adds it. Most of what I would still class as AI-as-an-accessory, the same lesson with a model stapled to the side, adds it. The rare exception earns its place by quietly taking work away. Very few right now clear that very high bar.

The Measured Response

So the pushback is half right, and the better half deserves saying plainly. The diagnosis is correct, and “skip crucial steps” is the most precise thing a politician has said about this in recent months, which makes Norway more interesting than Washington. Last week’s gatekeeper looked purely obstructive; this week’s has the diagnosis right.

The instrument is still wrong, though. Coming into Christchurch, the Southern Alps looked deceptively simple through the window (see below), a clean ridgeline you could imagine strolling across before lunch (well if you are Liberty Munson anyway!); you only meet the terrain on the ground. A ban has that quality from altitude. From the height of a press release it is a clean line through the problem, AI off, learning protected, job done. The climb is the lesson we still have to design and the assessment we need to create the morning the ban lifts, handed straight back to a fourteen-year-old and a phone it could never reach.

In edition 6 I argued the safest infrastructure for assessment is the kind no one can switch off; the same logic runs upstream into learning. The safest lesson is the one a ban never has to protect, because the thinking is built into the task and cannot be offloaded without the task collapsing. That is the re-innovation question. Not how we ban the forklift, but how we design the lesson, and the assessment, that makes the forklift useless. Build that, and a government’s off switch is a backstop, not the whole strategy. Fail to, and we are waiting for the next parliament to flick a switch we let someone else install.

That’s the Measured Response.

Over to you…

Here’s the homework (for the comments section):

A ban buys a few years. When it lifts at 14, what is the one thing you would build into the lesson, or the assessment, that would make a student want to take the crucial steps rather than skip them?

And a confession, to close the loop. A fortnight ago I said I would rather be wrong about governments, not technology, being the brake on AI in our sector. Two interventions later, Washington and now Oslo, I am still waiting to be.

Until next week, from the other side of the world…

Paul

Gabriel Demombynes’ World Bank blog, with the full figures from the China study, is here: https://blogs.worldbank.org/en/investinpeople/a-warning-shot-for-human-capital–evidence-of-an-ai-learning-pen

Jenny Anderson’s article for The Atlantic is here: https://www.theatlantic.com/ideas/2026/06/ai-tutor-education-human-investment/687678/