Measuring who it gets wrong
A leaf app says whether a leaf is healthy or sick. The app learned from photos of one garden.
- The app learned from garden photos only.
- Test on 8 garden leaves: all 8 right.
- Test on 10 forest leaves: only 6 right.
- Add forest photos and test again: 9 of 10.
The rule
Test each group on its own. A good total can hide a group it fails.
The app has one rule. It measures shine, a number it reads from the photo. Healthy leaves shine more than sick ones. The app learned the average shine of healthy and of sick leaves. In garden photos, healthy leaves average 78. Sick leaves average 44. It puts a line halfway between the two. (78 + 44) ÷ 2 = 61. A leaf that shines 61 or more is called healthy.
Forest leaves are duller. Some healthy forest leaves shine less than 61. So the app calls them sick.
All 18 test leaves together: 14 right. That total looks high. It hides the forest score of 6 of 10.
Bias means results that are unfair to one group. It often starts with the examples the AI learned from.
| Learned from | Garden score | Forest score |
|---|---|---|
| Garden photos only | 8 of 8 | 6 of 10 |
| Garden and forest photos | 8 of 8 | 9 of 10 |
More examples helped. With forest photos, healthy leaves average about 67 and sick leaves 37. (67 + 37) ÷ 2 = 52. So the line moved down to about 52. The app still gets one forest leaf wrong.
Switch on the forest photos. Watch both scores.
Loading lab…