Family AI Experiment: Train a Paper “Model” to Sort Snacks (Ages 5–10)

What You’re Building: A “Paper Model” That Learns to Sort Snacks

If your child has ever asked, “How does AI know what something is?”, this weekend project is a perfect answer—without screens, apps, or complicated setup.

In this family STEM activity, you’ll train a paper “model” (really: a simple set of rules you create together) to classify snacks into groups like:

• Crunchy vs. chewy
• Sweet vs. salty
• “Healthy-ish” vs. “treat” (your definitions!)

This is a simple classification activity for children that mirrors how many real machine learning systems work:

1. Pick labels (the categories)
2. Collect examples (training data)
3. Make a model (rules your model uses)
4. Test it (new snacks)
5. Fix mistakes (improve the model)

Kids ages 5–10 can do this with help. Younger kids enjoy the sorting and “model boss” role; older kids can track accuracy and redesign the rules.

Why snacks? They’re familiar, varied, and surprisingly “data-rich” (texture, shape, packaging, ingredients). It becomes one of those AI experiments for kids at home that feels like play—but teaches real ideas.

Materials + Setup (10 Minutes)

Keep it low-stress: you can do this with whatever’s already in your pantry.

You’ll need:

• 10–20 snack items (or pictures of snacks cut from flyers)
• Sticky notes or index cards
• A marker
• Paper for a “model card” (one sheet)
• Optional: small bowls/plates for category bins

Quick tip for families with allergies or limited snacks: Use toy foods, printed pictures, or drawings. The learning is the same.

Choose your categories (your “labels”)

Pick two categories to start—that keeps it clear and avoids decision fatigue.

Good starter pairs for ages 5–10:

• Crunchy / Not Crunchy
• Sweet / Not Sweet
• Has Chocolate / No Chocolate
• Bigger than a grape / Smaller than a grape (great for younger kids)

Write each label on a sticky note and place them as two “bins” on the table.

Pick training snacks vs. testing snacks

To feel like real machine learning, you want “data” your model has seen (training) and new stuff (test).

• Choose 6–10 snacks for training
• Set aside 4–8 snacks for testing

Try to include a few “tricky” ones (like granola bars that are sweet but also “healthy-ish”). Those are where the best conversations happen.

Train the Model: Turn Observations into Rules

Now you’ll build your “paper model.” This is the heart of the machine learning activity for kids.

Step 1: Human labeling (create the answer key)

Together, sort the training snacks into the two bins.

As you sort, ask:

• “Why did you put that there?”
• “What did you notice first?”
• “Could someone disagree? Why?”

This is how kids start thinking like ML designers: labels aren’t magic—they’re decisions.

Step 2: Extract “features” (what the model looks at)

Tell your child: “A model can’t taste. It needs clues.”

Have them pick 2–3 features that help decide the category. Examples:

• Texture clue: “Looks bumpy” / “Looks smooth”
• Ingredient clue: “Has chocolate chips”
• Sound clue: “Makes a crunch when you tap it”
• Shape clue: “Long bar” / “Round”

Write these features on the model card.

Step 3: Write your paper model (simple decision rules)

Create rules like a “choose your own adventure.” For example, for Crunchy / Not Crunchy:

• If it’s a chip or cracker, put it in Crunchy.
• If it’s a gummy, put it in Not Crunchy.
• If it’s a cookie, tap it gently. If it makes a loud sound, Crunchy.

Keep it short. If the rules get too long, the model becomes hard to use (a real AI problem!).

Here’s a ready-to-copy template:

Model Version	Labels (2)	Features We Use	Rules (Kid-Friendly)	Notes
v1	____ / ____	1) ____ 2) ____ 3) ____	1) If ____ then ____ 2) If ____ then ____ 3) Otherwise ____	What seems confusing?

Test Day: Measure Accuracy and Find “Edge Cases”

Now bring out the testing snacks—the ones your model hasn’t “seen.” Your child becomes the model operator.

How to run the test

For each test snack:

• The child uses the paper rules only (no “I just know!”)
• Place it in a bin
• Then the family checks: “Do we agree with that label?”

This is a great moment to explain:

• Prediction: what the model says
• Ground truth: what the family agrees is the right answer

Track results (simple, motivating data)

Use this table to make the results feel real and measurable.

Test Snack	Model’s Guess	Family’s Label	Correct?	Why it was tricky
______	______	______	✅ / ❌	______
______	______	______	✅ / ❌	______
______	______	______	✅ / ❌	______

Then calculate accuracy:

• Accuracy = (Correct guesses ÷ Total tests) × 100

For ages 5–7, you can skip the math and just count “wins.” For ages 8–10, let them do the calculation.

Look for the fun failures (that’s where learning happens)

When the model is wrong, don’t “fix” it immediately. Ask:

• “Which rule caused the mistake?”
• “What feature did we ignore?”
• “Is this snack an in-between snack?”

These wrong answers are called edge cases—examples that don’t fit neatly.

Improve It Together: Bias, Better Data, and Model Versions

Real machine learning isn’t about making a perfect first model. It’s about improving it responsibly.

Upgrade 1: Add better training examples

If your model keeps confusing two snacks, you might need more examples.

Try:

• Add 2–3 new training snacks similar to the confusing one
• Re-label them together
• Update the rules

This demonstrates a key ML idea: models learn from the data you give them.

Upgrade 2: Watch for “biased” rules

Sometimes kids create rules that are easy but unfair—like:

• “If the wrapper is green, it’s healthy.”

That might work on your pantry… until you buy different brands. This is a gentle way to introduce bias:

• A model can accidentally learn shortcuts that don’t generalize.

Ask:

• “Would this rule work at Grandma’s house?”
• “What if the same snack came in a different wrapper?”

Upgrade 3: Make version changes (v1 → v2)

Write a new model card:

• v1 rules
• v2 rules (after improvements)

Then test again and compare accuracy. Kids love seeing “leveling up.”

Optional challenge: Create a 3-class model

Once two categories feel easy, go for three:

• Crunchy / Chewy / Melts
• Sweet / Salty / Sour

Remind them: adding classes often makes the job harder—another real AI lesson.

Next Steps: Make This a Weekend AI Habit (No Tech Required)

If you want this to become a repeatable family STEM activity AI routine, here are practical ways to keep it going.

1) Try new “datasets” around the house

• Laundry: socks / shirts (or colors)
• Toys: building / pretend / art supplies
• Books: animals / space / magic
• Nature walk: leaves / rocks / flowers

2) Use the same experiment structure each time

• Pick labels
• Choose training vs. test items
• Write rules
• Test accuracy
• Improve model version

3) Keep a simple “Model Lab Notebook” On one page, record:

• Labels
• Features
• v1 accuracy
• v2 accuracy
• Biggest mistake (edge case)

4) One great question to end the activity Ask your child:

• “If we were building a real AI, what data would we need more of?”

That single question connects play to real-world AI thinking.

When you’re ready to take the next step, Intellect Council’s interactive lessons help kids build the same ideas (classification, data, training/testing) in age-appropriate coding and AI projects—while keeping the focus on curiosity and confidence.