Google Teachable Machine: Train Your Own AI in 10 Minutes (Step-by-Step)

I'm Parikshet. My favourite AI project of all time — the one I recommend to every kid who asks me "how do I actually do AI, not just read about it?" — is Google Teachable Machine. In 10 minutes, with no code and no downloads, you train a real neural network. Here is exactly how to do it.

What You Are Actually Doing

Teachable Machine is not a toy. It trains a real convolutional neural network (the same type used in professional image recognition systems) entirely in your browser. The model you build uses MobileNet — a lightweight but genuine neural network architecture designed by Google for efficient image classification.

When you show it examples of "thumbs up," you are giving it training data. When it learns to distinguish "thumbs up" from "thumbs down," it is doing the same thing that facial recognition, self-driving car vision, and medical image AI all do — just at a much smaller scale.

Step-by-Step: Build a Gesture Recogniser

Step 1: Go to teachablemachine.withgoogle.com. No account needed. Click "Get Started" → "Image Project" → "Standard image model."

Step 2: Name your classes. You will see "Class 1" and "Class 2." Rename them — "Thumbs Up" and "Thumbs Down" for this example. Click the pencil icon to rename.

Step 3: Record training examples. Click "Webcam" under "Thumbs Up." Hold your thumbs-up gesture toward the camera. Click "Hold to Record" and record about 100 images — move your hand slightly, change distance, change angle. Variety in training data makes better models. Repeat for "Thumbs Down."

Step 4: Train. Click the green "Train Model" button. Watch the accuracy rise in the training panel. Takes 30–60 seconds. No connection to external servers — this is running on your computer's graphics processor.

Step 5: Test. The preview panel goes live. Show each gesture and watch the confidence bars. If one class is weak, add more training examples for it and retrain.

Project Ideas to Try

Rock-paper-scissors detector — three classes, real-time gameplay.
Emotion detector — train on your own happy/sad/surprised faces.
Plant identifier — take photos of 3–4 plants in your garden and train a classifier.
Golf grip checker — my own project: I trained it on five different golf grip positions and used it to check my technique without needing a coach present.
Sign language letters — train on individual letters of the alphabet, then combine into a simple communicator.

What to Do After

Once your model works, export it:

Scratch: Use the ML4Kids Scratch extension to import your Teachable Machine model and build a game or interactive project around it.
Website: Export as TensorFlow.js and embed in a web page — the exported code is straightforward to use even with minimal JavaScript knowledge.
Python: Export as a Keras model and use it in a Python script. This is your first step into real machine learning development.

Every professional AI application starts with the same concept: collect labelled data, train a model, test it, improve it, deploy it. Teachable Machine gives you that full loop in 10 minutes. That is why I think it is the single best starting point in all of AI education for kids.