7 · Build & run the AI part (Teachable Machine)

Here's the fun part: building and running your AI experiment. For most capstones the easiest, safest tool is Google Teachable Machine — a free website where you train a real model in your browser by giving it examples, no coding needed. (If your project is AI-assisted analysis instead — like asking an AI helper to help organize data you collected — the same honesty and checking rules apply.)

Grown-up needed: Set up the tool with a parent, teacher, or mentor. They help you start it and stay nearby.

Train your model (the Teachable Machine recipe):

1. Make your classes. A class is one group you want the AI to tell apart — e.g. "ripe" and "unripe," or "clap" and "snap." Most starter projects use 2–3 classes.
2. Add examples to each class. This is the machine learning you learned in F3: the AI finds the pattern in your examples. Give it lots of examples per class — and make them varied (different angles, lighting, backgrounds) so it learns the real pattern, not a trick.
3. Train. Click train and let it learn the pattern from your examples.
4. Test on new things. This is the key step scientists never skip: test the model on examples it has never seen before. If you only test it on the same photos you trained it with, of course it looks perfect — that's not a fair test, it's just memorizing.

Keep it a fair experiment — remember Lesson 2:

• Change one thing at a time (your variable). Want to test if more examples help? Then change only the number of examples, nothing else.
• Write down everything as you go in your science notebook: how many examples per class, what you tested, what happened. Real data, recorded honestly.

The "did it really learn?" rule: a model that aces the training photos but flops on new ones didn't actually learn your pattern — it just memorized the examples (scientists call that "overfitting"). Always judge your model by how it does on new examples. That honesty is what makes it science.

You'll record your results next.

Think about it. Why must you test your model on photos it has never seen — not the same ones you trained it on? What would a "perfect" score on the training photos really mean?

Sources

• Google. Teachable Machine — train image, sound, and pose models in your browser, no coding. https://teachablemachine.withgoogle.com/
• MIT RAISE. Day of AI — free, hands-on activities for training models from your own examples. https://raise.mit.edu/