Rocket groans at the tally sheet. "Three plant photos called rocks! Let's start over."
Raven spreads out the three wrong cards. "Wait. What do you notice about these plants?"
"They are all dark and brown," Rocket says slowly. "Like dry leaves."
Nova floats closer. "What could you give the model," she asks, "so it learns those plants too?"
Building a model is a loop. Build it, test it, fix it, then test again.
The crew looked at the errors. Brown plants were called rocks.
So they added more labeled cards of brown, dry plants to the training set.
They also changed a rule: look for leaves before looking at color.
Then they tested again on test cards.
| Test | Right | Wrong | Cards tested |
|---|---|---|---|
| Test 1 | 7 | 3 | 10 |
| Test 2 | 9 | 1 | 10 |
Even a model that scores well can make mistakes.
Real AI is a tool people build. It can be wrong too.
That is why people keep testing and checking it.
A score like 9 out of 10 means it was wrong once in that test.
| Statement | True or false? |
|---|---|
| Testing again after a fix shows if the fix worked. | ? |
| A score of 9 out of 10 means the model is never wrong. | ? |
| Studying errors can show what training cards to add. | ? |
| Real AI is built by people and can be wrong. | ? |
Superb fixing! Tomorrow we look for testing in everyday life.