12 science experiments that show how AI learns

Train rock, paper, scissors with 10 examples each, test it, then retrain with 100 each. Kids measure the difference themselves. The cleanest demo of why data matters.

1. Hypothesis: more examples will make it more accurate.
2. Train with 10 images per class and test 20 times.
3. Retrain with 100 per class and test again.
4. Graph the scores.

Train 'cup' only in the kitchen and 'book' only in the bedroom. Then hold the cup in the bedroom. The model may say 'book' because it learned the room, not the object. A memorable lesson in hidden bias.

1. Record each object in a different room.
2. Train and test in the same rooms.
3. Swap rooms and test.
4. Fix it by recording both objects in both rooms.

Browse real doodles for one word and notice patterns, like most people drawing a shoe from the side. Then draw one a different way and see if Quick, Draw! recognizes it. Kids see that models learn what is common.

First train the bot with clear labels, then with fuzzy ones like 'angry fish'. Kids discover that when people label differently, the model inherits their opinions. A strong opener for talks about fairness.

Train a text model on kind and unkind sentences, then test sarcasm and slang. Kids see where language models struggle and why tone is hard for machines.

You hire people for a company, then let an algorithm learn from your choices. It copies and amplifies a bias. A short, eye-opening game about real-world harm.

Wave your arms in front of the webcam to conduct an orchestra. Test what breaks pose detection: dim light, baggy sleeves, standing too close. Kids learn about perception, one of AI's Big Ideas.

Watch for: Uses the webcam.

Add layers and neurons, press play and watch a small network learn to separate colored dots. Change one setting at a time and record what happens. A real lab for curious middle and high schoolers.

Type a sentence and see the probabilities a small language model gives for the next word. Slide the temperature to see answers get safer or wilder. It makes 'chatbots predict words' concrete.

Record the same birdsong from near and far, or with traffic noise, and see how the app's confidence changes. Kids learn that noisy input makes AI less sure.

Watch for: Uses mic and location.

The phone camera hunts for real objects matching an emoji. Try fooling it with drawings or photos of the object. Kids discover what a vision model actually looks for.

Watch for: Archived experiment; it may stop working. Uses the camera.

Face-analysis algorithms guess your age, attractiveness and more from the webcam, then explain how unreliable those guesses are. A powerful lesson on judgments AI should not make.

Watch for: It scores looks and body size, which can hurt. Preview it first and skip if your teen is sensitive about appearance.