🥭 → 🤖 “BANANA! 97%” → ❌

The phone holding up a big confident percentage next to completely the wrong answer.

🎒 What you need

your trained model from module 4, still open on the phone
the same objects
a dark corner, and a bright doorway

✋ Do this

Show it a mango in the same light as before. It gets it right. Fine.
Now walk to a shaded corner and show the same mango. Watch the numbers wobble.
Show it a mango on a different cloth or a different floor. Note what happens.
Show it something you never trained — a shoe, a cup, a goat if one wanders past. It will still pick a box, loudly.
Write down every way you managed to break it. Three ways is good. Five is excellent.
Now retrain, this time taking photos in shade AND sun AND on two backgrounds. Test again. Better?

💡 Why it works

The machine only knows the world you showed it. If every training photo was taken at noon on a blue cloth, then shade is a foreign country. This is why an app trained only on tidy photos from one place can fail badly on a real farm in harmattan dust — and why the people who build good ones go and collect messy photos on purpose.

🔥 Challenge

Be the tester, not the builder. Take a friend's model and break it in four ways in under three minutes. Then help them fix it.

📖 New words

confidencehow sure the machine says it is — which can be high and still wrong

training dataall the examples the machine learned from

🕵️

Bug Catcher

Tap when you have finished this module.