Stage 3: Test and Improve Your Model
Keep both tabs open all week. Open in a new tab — don’t use the buttons in this page to leave the course.
a model that is right more often than it was yesterday
how to find the AI's weak spots and fix them with better examples
a stronger model and a saved (new) URL
Show the room:
- Open your trained model from yesterday. Put your hand up to the camera. Show how the bars move.
- Move your hand to the side of the frame. Or tilt your head. Watch a bar drop and another one climb. "That's a weak spot."
- Add 5 new photos in the position that broke the model. Retrain. Show the bars get more confident.
- Say: "The AI was not wrong. It just hadn't seen this yet." Today is about teaching, not blaming.
The big idea
Yesterday you trained a model. The model works — sometimes. Today's lesson: find the times when it doesn't work, then fix them.
An AI is like a brand-new student. It learns from the examples you give it. If you only gave it Rock photos from one angle, it doesn't know Rock from any other angle. The AI is not broken — it just hasn't seen enough yet.
┌──────── Confidence bars ────────┐
│ Rock: ██████████████ 92% │ ← strong guess
│ Paper: ██ 8% │
│ Scissors: 0% │
│ Nothing: 0% │
└─────────────────────────────────┘
vs
┌──────── A weak spot ────────────┐
│ Rock: ██████ 51% │ ← wobbly guess
│ Paper: █████ 49% │ ← almost the same!
│ Scissors: 0% │
│ Nothing: 0% │
└─────────────────────────────────┘
A confident wrong guess is the AI's worst kind of mistake. A wobbly guess (close numbers, flickering bars) is the AI saying "I'm not sure" — which is honest. Today we hunt for both kinds and teach the AI better.
- confidence
- how sure the AI feels — shown by the percentage
- weak spot
- a situation where the AI gets it wrong or feels unsure
- retrain
- press Train Model again after adding new examples
- test
- try the model and look at what happens
Your model from Stage 2 should be trained. You should still see the Preview bars when you put your hand in front of the camera.
Build it
Step 1 — Hunt for weak spots
Put your hand in front of the camera and try all of these on purpose:
- Your normal Rock sign at normal distance.
- Your hand way up high in the frame.
- Your hand way down low.
- Your hand way to the left.
- Your hand way to the right.
- Your hand pulled close to the camera.
- Your hand far away.
- The room lights on. The room lights off (or partly off).
- Your hand tilted at an angle.
- A friend's hand making the same sign.
For each one, watch the Preview bars. Write down (on paper, or a notes app) at least TWO weak spots — places where the AI got it wrong, or where the bars flickered between two answers.
Step 2 — Pick your two worst weak spots
Look at your list. Pick the two that broke the model the most.
A good first weak spot is usually one of these:
- "My hand close to the camera looks like Nothing."
- "Rock and Nothing get mixed up when my hand is at the edge."
- "Scissors becomes Paper when I turn my hand sideways."
Each weak spot needs about 5 new photos taken in that exact bad situation.
Step 3 — Add 5 photos per weak spot
Go to the class that was getting it wrong. Click the webcam button.
For each weak spot:
- Recreate the bad situation. (Hand at edge of frame, hand close to camera, etc.)
- Make the correct hand sign for the class you're in.
- Hold Hold to Record for about 2 seconds. Move slightly while you hold.
- You should see 5 or so new photos appear.
Repeat for the other weak spot in its class.
Step 4 — Retrain and re-test
Click Train Model again. The wheel spins. Wait 30 seconds.
Now go back to the original weak spots and try them again. The bars should be more confident now. If they aren't, add 5 more photos to that class and retrain.
When at least 3 of your 4 hand signs work in the trickier situations, you're done.
Step 5 — Save the new URL
Click Export Model again. Click Upload my model. Get the new URL.
https://teachablemachine.withgoogle.com/models/AbC123xyz/
Important: the URL might be different from yesterday's. Replace your saved link with the new one.
Understand it
The number on each bar is confidence. It looks like a grade — but it isn't. Confidence is the AI's own opinion about its guess. A 95% bar means the AI is almost sure. A 55% bar means it's only a little more sure than flipping a coin.
The most dangerous AI mistake is a confident wrong answer. If your model is 98% sure your scissors are paper, the model thinks it's right. You'd need to tell it otherwise. That's why we test on purpose — to find those moments before they show up in the game.
The reason adding new photos to the right class fixes weak spots is simple. Imagine the AI is looking at your photo album. Yesterday's "Rock" album had 15 photos, all of your hand in the same spot. Today you added 5 more — same Rock sign, but in different spots. Tomorrow when it sees Rock in a new spot, its album now has examples close to that. It's not a fix, exactly — it's teaching.
There's a rule professional AI builders follow: the AI is only as fair as your photos. If you only train with one person's hands, the model may struggle with another person's hands. Stage 3 is where you find that out — and where you fix it by adding photos from more people, more lighting, more situations.
Try this
Try this
Three short experiments. Predict before you run, then test your guess.
Pick a hand sign and a brand-new situation you have not tested yet. Predict: will the bars be confident, wobbly, or just plain wrong? Try it. Were you right?
With your model running, ask a friend or coach to make the same hand signs at the camera. Do the bars look the same? If a different hand confuses the model, you've found one of the most important weak spots — and one of the most important fixes.
Stage 4 puts your model into RAISE Playground so the game can use it. Look at your weak spots one more time. Which weak spot would be the most embarrassing if it showed up while your parents are watching your demo on Friday?
Test your stage
- You wrote down at least 2 weak spots for your model.
- You added 5 new photos to each class that was getting it wrong.
- You retrained the model.
- At least 3 of the 4 hand signs work confidently in normal and the previously-broken situations.
- You exported a new URL and saved it (replacing the old one).
- Design check. Hold each hand sign for 5 seconds. Did any of the bars flicker between two answers? Less flicker = stronger model.
If it breaks
- The model got worse after I retrained. That can happen if you accidentally took photos of the wrong sign for a class. Open each class and quickly scroll through its photos — every photo in "Rock" should show a rock sign, etc. Delete any photo that doesn't belong.
- One class is way better than the others. Probably has way more photos. Roughly equal counts give the most balanced model. Aim for ~20 photos per class.
- The Preview bars look perfect, but the model fails in real life. You're testing in the same spot you trained in. Try moving your laptop somewhere different — different lighting, different background — and retest.
- My URL won't upload. Same as Stage 2 — refresh, retrain, retry. If it stays broken, ask a coach to check the internet connection.
This is the hardest stage of the week for kids 7–9. They want the model to be perfect; they read a wrong guess as they did something wrong. Reframe constantly: "The AI just hadn't seen that yet. You're teaching it. That's what training means." Praise the finding of weak spots, not just the fixing.
This is a 90-minute stage. Plan a movement break around the 45-minute mark. Photo capture is repetitive, and 7-year-olds will start drooping if you don't break it up.
The "test with a friend" stretch is genuinely valuable and worth pushing the room into if there's time — it's the first time campers experience the fairness dimension of AI in a concrete way. If a kid's model fails for another camper, that's an important moment to pause and discuss as a group.
Before campers move on to Stage 4, walk past every laptop and confirm a fresh URL was exported and saved. Stage 4 will not work without it.