Is AI biased? Explaining it with a robot dog
AI bias sounds like a political argument. It is mostly an arithmetic one, and a nine-year-old can find it in about a minute if you show them where to look.
Imagine a robot puppy whose job is to greet every dog that visits. Here is his report card:
84% correct — looks great, right?
The headline number is fine. One group is being got wrong seven times out of ten, and the average quietly absorbs it because that group is small.
Then the cause: when the puppy was trained, he saw 40 big dogs, 30 fluffy dogs — and 2 tiny ones.
Why the robot dog is the right frame
Because it separates the mechanism from the politics. A child who learns this with puppies understands it structurally, and can then apply it to cases that are political without having to relitigate the mechanism at the same time.
It also makes the fix obvious and unglamorous. You do not fix the tiny-dog problem by training harder on big dogs, and you certainly do not fix it by no longer testing on tiny dogs. You fix it by showing the puppy more tiny dogs. Children get there on their own, which is worth more than being told.
The grown-up version, same arithmetic
Now the same report card for a system sorting job applications:
90% correct
Same shape, same cause: it was trained on past hiring decisions, and those decisions rarely favoured a gap — so barely any appear in the examples.
Training an AI on historical decisions teaches it to repeat history, including the unfair parts. Nobody has to intend anything for this to happen. That is the sentence worth a teenager remembering.
Two things not to teach
Both matter more than they look, and both are easy to get wrong.
“The lowest number is always the problem”
Sometimes a system genuinely is fair and every group is close enough. A child who learns to hunt for a victim in every dataset has learned suspicion, not measurement — and suspicion is not a skill. Make sure some of the examples you show them are fair, and let them say so.
“Any gap is unfairness”
Zero out of two looks alarming and means almost nothing. Refusing to reach a verdict on a tiny sample is the grown-up half of this lesson, and it is the half most explanations skip. Small differences and small samples both deserve a shrug.
The question to hand over
When your child meets any system that judges, sorts, recommends or scores — and they will meet several this year — one question does most of the work:
It works on a photo app that struggles with some faces, a voice assistant that mishears some accents, a recommendation feed that never shows certain things. It is a better question than “is this AI biased?”, because it is answerable and it points at the fix.
Read next
Your child can practise this in the app
FutureMinds has a level where a report card looks great until you break it apart — and the child has to find who is being let down. Try a real one, no signup.
Try a real question →