Watch a child show off the photo app on a phone. It has sorted the pictures into little albums, and the child points, grinning: “See? It knows that’s a dog.”
It’s such a small sentence. It is also, we’ve come to think, the single most important thing to get straight about AI — and almost everyone, including adults building real systems, gets it slightly wrong in exactly that way.
The app does not know it’s a dog.
Here is the whole of what the app does. It turns the picture into a long list of numbers. It does arithmetic on those numbers. It compares the result to a line it worked out earlier from lots of example photos. Then it reports which side of the line this photo landed on. That’s it. There is no moment, anywhere in that process, where the machine understands what a dog is. It has never met a dog. It has no picture of one in its head, because it has no head. It matched a pattern of numbers and reported a side.
So when a child says “it knows,” they’ve done something very human and very misleading: they’ve slipped a mind into a machine that only has arithmetic.
Once you notice this, you start hearing it everywhere — and the mistake almost always rides in on a single word, a verb that belongs to people. The app knew it was a dog. The model decided that detail was important. The AI understood the question. It preferred the brighter photo. The system was sure. Every one of those verbs is borrowed from a mind, and every one is doing the same quiet damage: it hides the fact that a machine cannot know, decide, understand, prefer, or be sure of anything. It can only do sums.
This matters more than it looks. A machine that “knows” is something you trust. A machine that “did arithmetic and reported a side” is something you check. Same machine. Completely different relationship to it — and the difference lives entirely in the verb.
Here is the reflex we’d want every child (and honestly every adult) to have. When a minded verb gets attached to a machine, do two things.
First, swap the verb for what literally happened. “The app knew it was a dog” becomes “the app did arithmetic on the picture’s numbers and matched a pattern.” Less magical, much more accurate — and notice how the awe drains out and the checkability floods in.
Second — and this is the real prize — ask who actually decided. Because the agency didn’t vanish. It moved. Somebody chose which photos to train it on. Somebody chose what counts as “a dog.” Somebody read the answer and acted on it. The machine has no opinions, but the people around it had plenty, and every one of those choices is now baked into what the machine does. So the full move is: name the minded verb, replace it with the arithmetic, then point at the human whose choice is hiding behind the machine’s confidence.
We write books that teach children about AI, and of every wrong idea we’ve collected for those books, this one is the giant — the most common, and the root of most of the others. A child who thinks the machine “knows” will trust a confident wrong answer. A child who thinks the machine “decided to be unfair” will hunt for a villain in the code, when the unfairness actually came from a lopsided pile of examples a person chose. Get this one straight, and a surprising amount of the subject falls into place.
The machine is not a small mind. It is arithmetic, fast and cold, doing exactly what someone set it up to do. The app didn’t know it was a dog. The child did. The app just did the sums nobody sees.