The Idea
Machine learning works by showing an AI system many examples of something until it can recognize new examples on its own. This activity makes that abstract process completely physical.
Your child is the engineer. They create the training data, oversee the training process, and test the final model. The parent is the model — starting with zero knowledge and building up the ability to classify from examples.
The parent is not allowed to use any prior knowledge about what cats look like. They must learn purely from the examples they’re shown. (Yes, this is silly. That’s the point.)
Part 1: Create the Training Data (15 minutes)
Step 1: The engineer chooses a target
The child chooses what the “AI” will learn to recognize. Good options:
- Cats (the classic)
- Their drawings of a specific family member
- Made-up creature vs. real animal
- A specific emoji vs. other emojis
The target should be something distinctive but not something the parent would obviously recognize from real life (remember: the parent has “never seen” this before).
Step 2: Draw 10 positive examples
The child draws 10 examples of the target thing. Each on its own piece of paper. Encourage variety — different sizes, positions, styles — because real AI training data needs variety to work well.
Label each one on the back: “CAT” (or whatever your target is). These are your positive training examples.
Step 3: Draw 10 negative examples
The child draws 10 things that are not the target. Dogs, houses, cars, fish — anything that’s clearly not the target. Label each one on the back: “NOT CAT.”
Now you have 20 training examples total.
Part 2: Training the Model (10 minutes)
Step 4: Begin training
The parent closes their eyes. The child places the 20 examples face-up in front of the parent one at a time, announcing each time: “This is a cat” or “This is not a cat.”
The parent looks carefully at each example as it’s labeled. No touching, no stacking — just looking and learning.
After all 20 examples are shown, training is complete.
Step 5: The parent describes what they learned
Before testing, the parent has to explain (in their own words) how they would recognize a cat: “A cat has… [blank]. A non-cat usually has… [blank].”
This mimics how ML engineers try to understand what features their model is using.
Part 3: Testing the Model (10 minutes)
Step 6: Create test data
While the parent closes their eyes again, the child draws 6 new examples — 3 cats, 3 non-cats — that the parent hasn’t seen. These are new, not copies of the training data.
Step 7: Run the test
Show the parent each new example. They say “cat” or “not cat.” The child records their answers.
Calculate accuracy: correct answers ÷ 6 = accuracy. A random guess would be 50% accurate. A good model should do better.
Step 8: Analyze the mistakes
For any examples the parent got wrong:
- “What made this one confusing?”
- “What was different about the examples in training vs. this test example?”
- “If we added more training examples, which kinds would help most?”
The AI Connection
After the activity, connect it explicitly to machine learning:
“What just happened is almost exactly how machine learning works:
- We collected training data (your 20 drawings)
- We labeled that data (cat / not cat)
- We trained the model (showed the parent all the labeled examples)
- The model learned features (what makes a cat a cat)
- We tested the model on new data it hadn’t seen
- We measured accuracy
- We thought about how to improve the training data”
The main difference: real AI systems don’t have a person learning — they have mathematical algorithms finding patterns in thousands or millions of examples. But the structure is the same.
Discuss overfitting: Ask, “What if all 10 of your cat drawings had ears pointing left? And then in the test, the cat’s ears pointed right?” The parent might not recognize it as a cat. Real AI systems have this exact problem — they can learn the specific details of training data rather than the general concept. That’s called overfitting.
Discuss data variety: “Which drawing style helped the parent most — very detailed or simple? What if all 10 cats looked exactly the same?” Real AI needs diverse training data for exactly this reason.
Variations
Harder target: Try a more abstract concept — “happy face” vs. “sad face,” or “indoor scene” vs. “outdoor scene.” This is harder to specify and harder to learn, which mirrors real ML challenges.
Switch roles: Let the child be the model, learning from the parent’s training data. Does the child learn faster or slower? Why?
Team training: Two children each create their own training data for the same target, then test on each other’s test data. Whose training set produced a more accurate model?