Plain-language guide and offline activity

Machine Learning for Kids: How Computers Learn from Examples

Machine learning is a way for computers to find patterns in examples and use those patterns to make a prediction about something new. A person chooses the goal, examples, labels, and tests. Because examples can be incomplete or unfair, families should check results, question confidence, protect private data, and keep an adult involved.

Designed for ages 7–14 with adult-guided use. Younger learners can sort objects and describe clues; older learners can compare training choices, test failures, and investigate bias.

By the AI Education for Kids Editorial Team · Materially reviewed August 2, 2026

The vocabulary that makes models understandable

From examples to a tested prediction

Examples: Items the model can learn from, such as cards showing leaves and flowers.
Labels: Names attached to examples, such as “leaf” or “flower.”
Features: Clues that may help, such as shape, edge, color, or number of petals.
Training: The process of finding patterns in the labeled examples.
Testing: Checking the learned pattern with examples the model did not train on.
Prediction: The model’s best classification or estimated answer for a new input.
Confidence: A score describing how strongly the model favors an answer, not a promise that it is correct.
Bias: A repeated unfair or incomplete result that can come from examples, labels, goals, or design choices.
Overfitting: Memorizing training examples so closely that the model struggles with genuinely new ones.

Reproducible and account-free

Offline sorting classifier activity

Materials

Twenty small household objects or hand-drawn object cards; two paper category signs; pencil; paper; and an opaque bag or envelope. Use ordinary, non-private objects.

Steps

Choose a question with two labels, such as rolls and does not roll.
Select 12 training items. Place each by a label and say which visible or physical features guided the choice.
Write a one-sentence rule using those features. This rule is your human-built classifier.
Ask another person to hide the remaining eight test items in the bag.
Reveal one test item at a time. Predict its label using only the written rule, then test whether it rolls.
Record correct and incorrect predictions. Do not rewrite the result after seeing the answer.
Change one training choice or feature rule, then repeat the same eight-item test.

Expected observations

A useful feature, such as curved surface, may work better than color.
A confidence-sounding prediction can still be wrong.
More varied training examples can improve the rule on unfamiliar items.
A rule that perfectly describes the training set may fail on test items.

Common failure modes

Label disagreement: people interpret a category differently. Define the label before training.
Shortcut feature: the rule uses color or size even though it does not cause rolling. Add varied examples.
Data leakage: the tester reveals the hidden item before the prediction. Keep test items concealed.
Overfitting: the rule names individual training objects. Rewrite it using features that transfer.

Reflection questions

Which feature produced the most reliable predictions, and what evidence supports that?
Which test item exposed a weakness in the rule?
Whose real-world examples might be missing if this were a dataset?
When should a person ignore or override this classifier?

Three learning setups

Supervised, unsupervised, and reinforcement learning

Supervised learning

A model learns from examples paired with labels or answers. Our rolling activity is supervised because every training item has a category.

Unsupervised learning

A model groups examples by patterns without receiving the intended category names. A person still interprets what the groups mean.

Reinforcement learning

An agent tries actions and receives a reward signal. Over many trials it learns which actions tend to improve the total reward.

No commercial AI account is needed to understand these ideas. Families can model all three with cards, sorting, board-game rewards, and paper records before choosing any online tool.

What kind of answer does the model make?

Discriminative and generative models

Discriminative: choose between possibilities

This kind of model learns a boundary or rule that helps tell categories apart. It might decide whether a new card is more likely a leaf or a flower.

Generative: create a new example

This kind of model learns patterns well enough to produce something new, such as a sentence or image. New does not mean true, fair, or safe, so a person must review the output.

Limits, privacy, and source guidance

Use models as tools, not authorities

Do not upload a child’s face, voice, location, schoolwork with identifying details, or other private data.
An adult should review a tool’s minimum age, terms, data retention, sharing defaults, and deletion controls.
Test across varied examples and inspect mistakes; a high confidence score is not proof.
Do not use a classroom model to make medical, legal, disciplinary, or other high-stakes decisions.

Google Teachable Machine — an official browser-based tool to explore model training; an adult should review privacy and age suitability first.
UNESCO AI Competency Framework for Students — primary guidance on human-centered, ethical, technical, and design competencies.

Want a guided first lesson?

The free class introduces AI, hallucinations, privacy, and responsible use with parent support.

See the Free Class