Build Your Own AI That Recognizes Objects: A Kids Workshop Using Google Teachable Machine

What you'll learn in this article:

• What Google Teachable Machine is and why it is the best tool for teaching kids how AI actually learns
• Full step-by-step instructions for training a 3-class image classifier
• How to use a free ChatGPT prompt to deepen the learning after the model is built
• What core AI concept the activity teaches and why it matters for the future
• A parent/educator sidebar with conversation starters and learning extensions

Every facial recognition system, every product defect detector on a factory line, every self-driving car's object recognition module works on the same fundamental principle: a model trains on labeled examples, learns the visual patterns associated with each label, and then makes predictions on new images based on those patterns. This is called supervised machine learning, and it is the core concept behind a large fraction of commercial AI systems deployed today.

A child aged 8 to 16 can build a working version of this in under 15 minutes using Google Teachable Machine -- a free, browser-based tool from Google that trains a real image classifier using webcam photos. No code. No account required. No cost. The child chooses 3 objects, records webcam photos of each, trains the model, and tests it in real time -- watching the confidence scores change as the model decides what it is looking at.

After the model is built, a single ChatGPT prompt turns the experience into a 5-minute conversation about why AI makes mistakes, what training data quality means, and what the child would need to do to make the model better. That conversation is the one that sticks.

What You Need

• A laptop or tablet with a built-in webcam (any computer made after 2015 will work)
• A browser (Chrome works best; Firefox, Safari, and Edge also work)
• 3 physical objects to classify (pencil, water bottle, phone; book, eraser, cup -- any 3 visually distinct objects)
• No account, no downloads, no cost

Full Activity Instructions

Part 1: Set Up the Project (5 minutes)

Step 1: Open a browser and go to teachablemachine.withgoogle.com

Step 2: Click "Get Started" on the homepage.

Step 3: On the project selection screen, click "Image Project."

Step 4: Select "Standard image model" (the default option).

Step 5: The screen now shows a workspace with two empty classes: "Class 1" and "Class 2." These are the categories the AI will learn to recognize.

Part 2: Name the Classes and Collect Training Images (5 minutes)

Step 6: Click the pencil icon next to "Class 1" and rename it to the name of your first object (e.g., "Pencil").

Step 7: Click the pencil icon next to "Class 2" and rename it to the name of your second object (e.g., "Water bottle").

Step 8: Click "Add a class" at the bottom and name the third class (e.g., "Phone").

Step 9: For each class, click "Webcam" and hold the object in front of the camera. Click and hold the blue "Hold to Record" button to take photos -- aim for at least 50 per class. Try different angles, distances, and lighting conditions.

Why 50 photos matters: Explain to the child that the AI can only learn from what it sees. If all the photos show the pencil from the same angle with the same lighting, the model will struggle to recognize it from a different angle. This is training data diversity -- the same reason AI systems that train on data from one country sometimes fail in another.

Part 3: Train the Model (2 minutes)

Step 10: Once all three classes have at least 50 photos each, click the blue "Train Model" button.

Step 11: Wait. Training typically takes 30-60 seconds. A progress bar shows the model learning.

What is happening during training: The model is running through all the photos thousands of times, adjusting its internal pattern-matching weights until it can distinguish the visual features that separate one class from another. This is the same process -- just much smaller -- that happens when large AI models train on billions of images.

Part 4: Test It Live (3 minutes)

Step 12: Click "Preview" (if not already showing). The webcam opens and the model starts making predictions in real time.

Step 13: Hold up each object and watch the confidence score bars.

Things to try:

• Hold the object from the angle the model was trained on -- it should score 90%+ confidence
• Flip the object or hold it at an unusual angle -- watch the confidence drop
• Hold two objects at the same time -- observe which one the model picks and how confident it is
• Move the object closer or farther -- watch how distance affects the score
• Cover part of the object -- notice when the model loses confidence

Each experiment is a concrete demonstration of a key AI concept: the model can only recognize what matches the patterns in its training data.

After the Activity: The ChatGPT Conversation

Once the model is built and tested, open ChatGPT (free, no account required at chatgpt.com) and paste the following prompt:

THE PROMPT

"I just trained an image classifier to recognize these 3 objects: [child fills in their object names]. The AI looks at pixels and learns patterns. Can you explain in simple terms why it sometimes gets confused when I hold up two objects at the same time? And what could I do to make it more accurate -- like take more photos, or choose more different objects, or try different lighting? Also, what real-world jobs or products use this same type of AI?"

What a good response covers:

• Why simultaneous objects confuse classifiers (the model was trained on single-object images, so multi-object inputs do not match any training pattern well)
• How to improve accuracy: more diverse training data, more varied lighting and angles, clearly different objects (high visual contrast between classes)
• Real-world applications: medical imaging (identifying cells), manufacturing quality control (spotting defects), retail inventory counting, wildlife monitoring, self-driving car object detection

This 5-minute ChatGPT conversation turns the hands-on activity into a conceptual framework the child can apply to any AI system they encounter.

Learning Extensions

For younger children (8-11):

Draw a picture of what the AI "sees" -- a grid of colored squares (pixels) -- and explain that the model learned to recognize which color patterns go with which object. This makes the abstract concept of pattern matching concrete.

For older children (12-16):

Ask: "What would happen if you trained the model on photos of someone's face and then showed it a photo of a different person? What ethical questions does that raise?" This connects the technical activity to real discussions about facial recognition, privacy, and bias in AI systems.

Extension activity:

Export the trained model from Teachable Machine (click "Export Model") and embed it in a simple p5.js sketch using the ml5.js library -- a free, beginner-friendly JavaScript library designed for creative AI projects. This optional extension introduces the child to the concept of deploying a model into a running application.

For Parents and Educators

Core AI concept this activity teaches: Supervised machine learning -- the process by which an AI model learns to classify new examples based on labeled training data. This is the foundational concept behind image recognition, spam detection, medical diagnosis AI, and most commercial AI classification systems.

Conversation Starters:

• "Your AI learned from the photos you gave it. What do you think happens when an AI learns from photos that only show one type of person, or one country, or one kind of environment? What might it get wrong?"
• "You noticed the model was less confident when you covered part of the object. What does that tell us about what the AI is actually 'looking at' -- is it looking at the whole object, or specific parts of it?"
• "If you were building an AI for a hospital that needed to recognize healthy versus sick cells, how would you make sure the training data was good enough to trust? Who would you want to check the AI's work?"

These questions connect the hands-on activity to broader discussions about AI bias, training data quality, and human oversight -- topics that will be relevant throughout the child's lifetime.

Action Steps

1. Open teachablemachine.withgoogle.com and start an Image Project. No account or download required. Works in Chrome on any laptop or tablet with a webcam.

2. Create 3 named classes and record at least 50 webcam photos of each object. Vary the angle, distance, and lighting while recording to build a more robust training set.

3. Click "Train Model" and observe the training progress bar. Explain to the child that the model is running through all the photos repeatedly to learn the patterns.

4. Test the model live and run at least 3 experiments. Try unusual angles, covering part of the object, and holding two objects at once. Discuss what each result means.

5. Run the ChatGPT prompt to connect the activity to the real world. Paste the verbatim prompt into free ChatGPT to extend the learning into real-world applications and improvement strategies.