How AI “Guesses” — and When It Is Wrong
When a program “recognises” a cat in a picture, it does not look the way you do. Today you will build your own tiny “guesser” with cards — and see how it can be fooled.
For parents: a one-page guide — dinner questions, a game without a screen and the age conditions of chatbots.
01Remember
Before we touch the program, think about three everyday questions. Then open the answer.
How do you tell a cat from a dog in a picture?
By the ears, the snout, the tail, the whiskers, the size… Each of these is a clue (a “feature”). You rarely go by one — you use several together.
Have you ever mistaken someone from far away?
Maybe they had the same jacket or the same walk. You decided from a few clues — and got it wrong. That is completely normal.
How do you know what a dog looks like?
You have seen many dogs and someone told you “that is a dog”. You learn from examples. Some programs learn in a similar way.
02Story
In 1957, the American scientist Frank Rosenblatt worked at the Cornell Aeronautical Laboratory in the USA. He wrote a program for the IBM 704 computer. The program learns from examples, and he called it the “perceptron”.
Then he wanted a real machine too — and built the Mark I Perceptron, made to recognise images. In one version, its “eye” was 400 light-sensitive cells arranged in a 20-by-20 square. That is 400 dots — far fewer than a phone picture.
The machine has been kept and is in a museum in the USA today. The program in this lesson is much simpler — but it also learns from examples.
03Worked example
Our program has learned from 10 cards: 5 cats and 5 dogs. Each card is drawn with three clues: ears, snout and whiskers. Here is a new card: pointy ears, short snout, whiskers. For each clue the program asks: “How many of the cards with this clue are cats?”
Ears: pointy
There are 4 cards with pointy ears in the training set. All 4 are cats. That is 4 out of 4 = 100% for cat.
Snout: short
There are 6 cards with a short snout, and 5 of them are cats. That is 5 out of 6 = 83%.
Whiskers: yes
There are 5 cards with whiskers, and 4 of them are cats. That is 4 out of 5 = 80%.
Average
(100 + 83 + 80) ÷ 3 ≈ 88% cat. The program's rule: above 60% — “cat”, below 40% — “dog”, in between — “I'm not sure”.
See? The program does not “think” or “see” like you. It only counts and takes an average. That is why it can get confused when the examples are few or all alike.
04Try it now
1 · Predict
These are the 10 cards the program learns from. Look at them carefully.
And this is a new card. It is drawn as a dog: pointy ears, short snout, no whiskers.
What do you think the program will say? Choose before you check.
2 · Run
Draw your own card by choosing the clues, or pick a ready-made one. Then press “Guess!”.
Ready-made cards:
The bar shows the estimate in per cent “cat”. The white marks are at 40% and 60%.
3 · Change: fix the training
The card from “Run” (with the clues you chose) can join the training set. You decide the label: cat or dog. Add a few cards and press “Guess!” again. How many cards does it take for the program to change its mind?
Notice: when you add more varied examples, the program changes. But even then it sometimes says “I'm not sure”. And the rabbit? It is neither a cat nor a dog, but the program has nowhere else to put it.
05No screen: “Guess by 3 clues”
You need: 10 slips of paper, a pencil and one more person.
On each slip draw an animal head with three clues: ears (pointy or floppy), snout (short or long), whiskers (yes or none). On the back write “cat” or “dog”.
Show the other person 5 slips with the labels. That is the “training”. Then take a new slip without showing the back and ask: “Cat or dog? Which clue helped you?”
Make one “tricky” card — for example a dog with pointy ears. What will the other person say? Why?
Show 3 more labelled slips. Try the tricky card again. Has the answer changed?
06Make your own
Invent your own “guesser” for two other things — for example an apple and a pear, or two kinds of leaves.
- Choose 3 clues that tell them apart.
- Make 10 cards with labels.
- Think of one “tricky” card where the program will go wrong. Write down why.
- Explain to someone at home how the program learns and why it should not be trusted blindly.
07What I learned
- A program learns from examples with labels and guesses by similarity to them.
- If the examples are few or all alike, it gets things wrong. More and more varied examples help.
- A percentage is an estimate, not certainty. AI can be wrong — check with a person.
You also sometimes guess from a few clues. What do you do to check?
Check yourself
1. The program says “cat” for a dog with pointy ears. Why?
2. How can such a mistake be reduced?
3. The program says “cat · 90%”. What does that mean?
It is yours when you can show someone a tricky card and explain why the program gets confused. No points and no rankings — just a skill.
08I have a question
Do you still have a question? Write it here. It goes to your parent, not to strangers. Your parent confirms their e-mail first.
The form is loading… If it does not appear, write your question on paper and show it to a parent or teacher.
·Sources
- Title: How AI “Guesses” — and When It Is Wrong (DB-IB-06) · Author: Children's Library, East Accelerator Foundation · Source: original content, 2026 · Rights: CC BY-SA 4.0.
- The story: facts from the article “Perceptron” on Wikipedia (read on 02.10.2026), retold in our own words.
- The cards and the classifier were drawn and written by us; nothing is copied.
- Licence: CC BY-SA 4.0 · Children's Library, East Accelerator Foundation, 2026. You may share and adapt this lesson if you credit the author and keep the same licence. Learning is free forever.