Romane had put the ketchup with her socks and a screwdriver. Her criterion: red stuff.
That's how the activity started — by accident, on a Tuesday morning, before I'd had time to make my coffee.
What AI actually does
When an AI "recognizes" something (a photo of a cat, a spam email, a disease on a scan), it does exactly what Romane did with the ketchup: it looks for common features and creates groups.
In technical terms, it's called classification. But the real story is the sorting criterion. Because changing the criterion completely changes the result — and therefore changes what the AI "sees."
The activity
What you need:
- About twenty small varied objects (coins, pebbles, leaves, cutlery, toys, bottle caps…)
- Large sheets of paper to mark out zones
- Markers to label the groups
- 40 minutes and a little chaos
How it works:
-
Dump everything on the table. Let the kids look without any instructions.
-
First question: "If you had to put all this away, how would you do it?" Let each child sort their own way, without stepping in.
-
Show that the criterion changes everything. Take the same objects and re-sort them by a different criterion (color, material, size, weight, "things that make noise"). Each time, the groups change completely.
-
The AI challenge. "An AI is a program trained to choose a criterion. But if it learns the wrong one, it gets things wrong. What would be a terrible criterion for sorting these objects?" (Romane answered: "Things that have already touched the floor.")
-
Swap roles. Romane picks a secret criterion and makes her groups. I try to guess. That's exactly what a researcher does when analyzing an AI.
What actually happened
Romane quickly understood that the criterion was the real power. She spent ten minutes inventing absurd categories and was delighted every time I couldn't figure them out.
Meryl participated in his own way. He picked up objects, turned them over, put them somewhere else. No apparent criterion — just exploration. At the moment we finally had a perfect sort, he put his elbow on the table and knocked everything over. We had to start again. Which, in hindsight, was a pretty good illustration of what's called "noise in the data."
At one point, Romane asked me: "Dad, does AI know when it's wrong?"
I didn't answer right away.
A question to end on
If we trained an AI to sort these objects using Romane's criterion (red stuff), would it be right or wrong about the ketchup?
Both. Depending on what we were actually trying to sort.