A sparrow flies. A pigeon flies. A parrot flies. So all birds can fly—right? One penguin is enough to challenge that rule.
Your child may already know the answer. The interesting question is what someone could conclude if the only bird examples they had seen were flying ones. This paper-card activity makes that gap visible and opens a conversation about how examples can shape predictions.
Why explore this today?
Common Sense Education's Digital Citizenship Week runs from 19–23 October 2026. Its “How Does AI Learn?” activity explores how training data affects an image recognition model. As the week closes, BAIO is taking that broad idea to the family table with an original, screen-free exercise for children in Classes 3–8.
This is a discussion game, not a real AI model or an official Common Sense lesson. It illustrates one possible source of mistakes: examples that leave out important cases.
Prepare twelve animal cards
Use twelve small pieces of paper. Write one animal name on each; drawings are optional. An adult can read the cards aloud and explain unfamiliar animals. No photographs of children, online accounts or AI tools are needed.
First set: sparrow, pigeon, parrot, dog, cat and cow. On the back of the first three, write “bird”; on the back of the other three, write “not a bird”. These are the examples you will start with.
Second set: penguin, ostrich and bat. Label penguin and ostrich “bird”, and bat “not a bird”. Keep this set aside until the first rule has been discussed.
Final set: eagle, kiwi and butterfly. Label eagle and kiwi “bird”, and butterfly “not a bird”. Explain that kiwi here means the flightless bird, not the fruit. Keep these cards separate for a later check.
Round one: find a rule
Show the first six cards and their labels. Ask, “What rule might help us sort these into birds and other animals?” Write down your child's suggestion. There can be more than one rule that fits this small set.
If they suggest “birds can fly”, explore that rule. If they already suggest feathers or mention penguins, welcome the observation. Do not ask them to give a wrong answer. Instead, have the adult propose a deliberately limited rule: “If an animal can fly, call it a bird; otherwise, call it not a bird.”
Try the adult's rule on the first six cards. It puts all six in their labelled groups. Now ask, “Does getting every example here right prove that the rule will work for every animal?” Record the prediction before opening the next set.
Round two: reveal the missing cases
Bring out the penguin, ostrich and bat cards. Apply the flying rule exactly as written. It would reject the two flightless birds and accept the bat, even though bats are mammals. The rule fails on all three of these deliberately chosen challenge cards.
This is a result of our paper game, not a measured error rate for an AI system. We chose these cards to expose the weakness in a rule that looked successful earlier.
Ask your child which examples were missing from the first set. Flightless birds show that flying is not necessary for being a bird. A flying animal that is not a bird shows that flying is not sufficient either. Younger children can explain this in their own words without learning those terms.
Round three: revise, then check again
Discuss a better clue. For these animal cards, having feathers distinguishes the birds from the other animals. Write a revised rule, then use the final three cards to check it: an eagle and a kiwi belong in the bird group; a butterfly does not.
Keep the final set out of the rule-building discussion until this point. If you change the rule after seeing it, find fresh examples for another check. Explain the purpose: we want to see whether an idea works beyond the cases we used to build it.
This little check does not certify a real image recognition system. A written fact about feathers is different from a computer detecting feathers in a photograph, where lighting, background and what is visible can vary.
What does this tell us about AI?
In supervised machine learning, a system learns from examples paired with target answers, often called labels. Those examples are training data. A model that works on familiar examples still needs checking on new ones. Google’s Machine Learning Crash Course explains that strong training performance can coexist with poor performance on unseen data, and that unrepresentative examples can contribute to this problem.
Our human-written flying rule is an analogy for limited evidence, not a demonstration of how every AI learns. Real models learn patterns through computation; they do not necessarily produce a neat rule we can read. This exercise also does not explain every chatbot error.
More copies of the same six cards would still leave the missing cases out. For our activity, adding different, relevant examples changes what we can notice. In a real project, the useful next question is which situations the system needs to handle and whether its examples and checks cover them.
Finish with three questions
Ask your child: “Which example changed your mind?” Then ask, “What would you test next?” Finally: “What can you now explain about this rule that you could not explain at the start?” Praise a clear reason or a thoughtful revision rather than a perfect first guess.
For Classes 3–5, focus on sorting and explaining. For Classes 6–8, introduce training examples, labels and separate checking examples. Let older children invent a different sorting task and look for a case that breaks their first rule.
Turn an AI mistake into a better question
Our earlier family challenge asked children to spot AI mistakes. This activity adds a follow-up: what examples or checks might be missing? Our recommendation game offers another way to explore patterns and choices; both articles are linked below.
That habit fits BAIO's focus on reasoning: propose an idea, test it and explain a revision. The useful takeaway is simple enough to share at dinner: getting the practice examples right is a beginning. Try a new example before trusting the rule.
Key Learning Self-Check
Test your understanding of the essential insights from this guide:
Did you find this guide insightful?
Click an emoji to share your reaction with the BAIO editorial team
Frequently Asked Questions
Does this card game train a real AI model?
No. The card game uses a rule written by people to illustrate how limited examples can lead to mistakes. It does not run machine learning software or measure the performance of an AI system.
What if my child already knows that penguins cannot fly?
That is useful knowledge. In the card game, let the adult propose the flying rule and ask your child to test it. The goal is to examine a rule, not to trick a child into a wrong answer.
Does adding more examples fix every AI mistake?
No. In this card game, different relevant examples reveal a missing case. Real AI errors can have several causes, and adding data alone does not guarantee reliable results. Testing on new examples remains important.