The AI said OK. Nothing moved.

A child said “set the fan to low.” The AI picked the command FAN 30. The board answered OK FAN 30. And nothing moved — because no fan was wired to that board. The sentence was reasonable, the command was real, the machine said OK, and every word on that screen was useless. The only person in the room who noticed was the child watching the log.

That was Saturday 5 September, in our studio in Bahria Town, with nine kids around one bench. It was not an accident. We had wired the fan station without a fan on purpose, because that five seconds teaches something no amount of prompt-writing practice does: an AI will answer you confidently when it has nothing behind the answer, and the skill worth having is noticing.

First they learn what the machine actually speaks

Most “AI for kids” sessions hand a child a chat box and call the output a project. We run the day the other way round. Before anyone was allowed to say a sentence out loud, the kids typed raw commands into an Arduino by hand — twenty-two of them, in the machine's own wire format:

• BARRIER OPEN — lift the cardboard car barrier on the desk.
• HOLD 5 — open it, wait five seconds, close it again.
• SAFE — ask the gas sensor whether the air is safe, and get a number back.
• BLINK 5, MORSE SOS, DIST, TUNE TWINKLE — and fifteen more.

Only then did we turn the microphone on. And notice what that does to the AI's job: it is now small, specific and checkable. Take a sentence a child says in their own words, pick one command out of a list the child already knows, and show its work. A kid who has typed BARRIER OPEN with their own fingers can tell in a glance whether the AI chose right. A kid who has only ever chatted cannot — for them the box is magic, and magic can't be audited.

Three boxes, and no magic left

So the screen shows every sentence three times over. What you said. What the AI chose. What the board sent back. Saturday's real ones, off the station screens:

1. 01

   “open the barrier”

   The AI chose BARRIER OPEN. The board replied OK BARRIER OPEN, the servo turned, and a lolly stick taped to a cardboard box went up. Nothing on the screen counted until that stick moved.

2. 02

   “keep the barrier open for 5 seconds”

   The AI chose HOLD 5 — not two commands, one. The board opened the gate, waited, and closed it. A child who knew the list could see that the AI had picked the tidier of two right answers.

3. 03

   “is the air safe”

   The AI chose SAFE. The board answered OK SAFE YES, and a reading. That last box is the part that matters: the answer came from a sensor in the room, not from the model.

That middle box is the whole design. It is where a guess stops being a guess — and it is also where you can see the guess go wrong while the system still says OK.

Then we broke it on purpose

Which brings us back to the fan. FAN 30 is a genuine command; the firmware runs it; the board acknowledged it. The failure was one layer below everything anyone was looking at — a component that wasn't there. No error appeared, because from the software's point of view nothing had gone wrong.

We think an OK with nothing behind it is worse than an error. An error teaches a child to look. An OK teaches them to trust the words on the screen, which is exactly the habit we are trying to prevent. So our console is built to admit when the board did not answer, rather than smoothing it over — and it will refuse a command outright if the AI invents something plausible that the firmware has never heard of. The list of twenty-two is the leash. The AI can pick from it. It cannot make one up and have the room nod along.

At the other table, they were debugging their own games

The same principle ran the other half of the room, where kids were building browser games by directing an AI one sentence at a time — then playing them, finding what broke, and working out why. One is a cowboy dinosaur roping runaway dinos in a mine, with a boss who has a health bar and a name. The other is a slingshot against pottery kilns: two dozen levels, each one named, each scored out of three stars.

You can tell instantly that no AI wrote these on its own, because of the writing. The dinosaur game signs off with a line one of the kids put there — “Gate stays SHUT. Nobody opens it. Ever.” — on the same afternoon the table next to him was opening a gate by asking it nicely. That is a child's joke, and it is worth more than the code.

The AI drafted. It also broke things, repeatedly, the way it always does. Working out why it broke is not the annoying part of the session — it is the session. It is the skill that survives every new model release, because it is the only part of the loop the model isn't doing.

“But doesn't AI make kids lazy?”

Yes. The version of it most children meet absolutely does, and a parent who says so is describing something real: paste the homework in, paste the answer out, learn nothing, and lose the ability to tell whether the answer was any good. We are not going to argue with that, because it is true.

The difference is not enthusiasm, it is structure. A child who has typed the command by hand can audit the AI's choice. A machine on the desk decides who was right, rather than the model marking its own homework. And the failures are put in the child's path deliberately, early, while an adult is standing there — instead of arriving years later on something that matters.

There is a real gain on the other side of it, too. A fourteen-year-old with a good idea used to spend months on syntax before building anything they cared about. Now the idea is reachable in an afternoon — provided they can read what came back, change it, and catch it when it is wrong. That last clause is the entire course.

Try the thing itself, tonight, for free

You do not have to take any of this on trust, and you should not have to. The console the kids used on Saturday is on this site, free, with no login: type or say a sentence and watch it become a command. You do not even need a board — without one, the first two boxes still fill in, so you can see exactly what it would have sent, and the third box tells you plainly that there is nothing connected. Which is, of course, the whole point of this article.

One honest catch: it drives the board over Web Serial, which means a laptop running Chrome or Edge. Phones and tablets do not have it, and Safari does not implement it. We would rather say that here than have you find out with a child waiting.

And if you would like the full write-up of the day — the rig, the real transcripts, both games running, and the honest notes on what we would change next time — it is all in the case study.