Your child’s school has put “AI” on the timetable. Maybe it’s a bit of Python, maybe a chatbot, maybe an app that sorts photos into folders. And you might be quietly wondering what, under all of it, she is actually meant to come away with.
Here is our honest answer, after building books for exactly this subject: not the coding.
The specific tool she learns this year will be gone in three. The app will vanish; the model will be old news before she finishes school. If the tool is the lesson, the lesson expires. So the question we kept asking while building was: which part won’t expire — what will still be working for her at twenty-five, whatever the technology looks like by then?
It comes down to two questions. Small ones — small enough that she can ask them without thinking, the way she looks both ways before crossing a road.
First: what was actually measured?
Almost every AI system is really an argument made from data. It was shown a pile of examples, and it generalises from them. So the first thing to ask about any confident output is what was in the pile. Say an app claims it can predict which students will do well. It did not measure “doing well” — there is no such number. It measured something standing in for it: last year’s marks, maybe attendance, maybe which school you came from. The gap between the thing you care about and the thing actually counted is where most quiet damage hides. A child who hears a confident answer and simply believes it is at the mercy of that gap. A child who asks “okay, but what did it actually measure?” can pry the gap open and look inside. Same child — one reflex different.
Second: compared to whom?
Almost every data claim is secretly a comparison, and the comparison group is usually invisible. “This is normal.” Normal compared to whom? “This face wasn’t recognised.” Recognised less well than whose faces — and why might that be? “Most students prefer this.” Which students were asked? The moment a child learns that every claim about “normal” or “best” or “most” is smuggling in a reference group — and that who is in that group decides the answer — she has a permanent defence against being quietly sorted by a number she didn’t choose.
That’s it. What was measured, and compared to whom. Two questions a twelve-year-old can hold — and a surprising number of adults building real systems forget to ask.
In our books, these two questions are not a chapter. They are a thread that runs through every year of the subject, coming back each time at a deeper level — a tally chart of favourite fruits in one year, the training set behind a face-recognition system a few years later. Same reflex, harder object. (We’ve written separately about why we structure everything this way — a habit is built by being used until it’s automatic, not by being introduced once and admired.)
So here is what “teaching a child AI” means to us. It does not mean producing a small engineer — most children will never build models, and that is completely fine. It means producing someone who, for the rest of her life, when a confident system or a confident person hands her a number, hears a quiet voice ask: what was actually measured here — and compared to whom? The Python will date. The apps will vanish. Those two questions will still be working for her at forty, in a hospital waiting room reading a risk score, in an argument about a statistic, in front of whatever AI exists by then.
That’s the cargo. The subject is just the vehicle that delivers it — and if the books do their job, she won’t even notice she’s been taught to think. Only that the magic stopped looking quite so much like magic.