A First Machine Learning Project for Kids—Without Personal Data

Choose categories that do not describe people

Start with two categories of child-created drawings, such as stars and hearts, spirals and zigzags, or sunny and rainy symbols. Avoid faces, voices, handwriting tied to names, school records, locations, or private messages.

The learning goal is to observe how examples affect a classifier. Personal data adds risk without improving that goal. UNICEF’s child-centred AI guidance emphasizes privacy, safety, fairness, transparency, and children’s well-being.

Build a small, balanced training set

Ask several people to draw each safe category on blank cards or in a local drawing tool. Use the same number of examples for both labels. Include reasonable variety: large and small shapes, thick and thin lines, and different drawing styles.

Explain that a label is the answer attached to an example. The model looks for patterns connecting examples to labels. It does not understand a heart the way a person does.

• Count the examples in each category
• Keep test drawings separate from training drawings
• Remove names or identifying marks
• Record what kinds of variety are included

Predict before testing

Before showing the model a new drawing, ask the learner to predict whether it will be easy or confusing and why. Test clear examples first, then rotate a shape, draw it smaller, overlap two shapes, or use a style missing from the training set.

Record the model’s guess, the expected label, and one possible reason for the result. A mistake is not proof that machine learning is useless. It is evidence about the examples and patterns the model had available.

Improve the data, then test again

If tiny stars are often misclassified, add several tiny stars to the training set. If one person’s drawings fail more often, add varied examples from that style. Retrain and repeat the same tests so the comparison is meaningful.

This is the central lesson: performance depends on the data, labels, design choices, and test cases selected by people.

Finish with a model card

Create a one-page model card written by the child. It should name the two categories, describe the training examples, list situations where the model worked, list situations where it struggled, and state that it should not be used to make decisions about people.

That final explanation turns a novelty demo into an AI literacy project. The learner can describe not only what the model guessed, but also why its limits matter.

Keep exploring

• Try seven more beginner AI projects
• Learn about bias and fairness
• Explore the AI Labs