In the story, the station has a pretend voice door. Rocket says "Open!" and it slides open.
Raven says "Open!" in her quieter voice. Nothing happens.
"It hears me fine," Rocket shrugs.
Raven frowns. "What do you notice about whose voices it learned from?"
Nova hums. "Who was in the room when the door was trained?" she asks.
The voice door in our story is pretend. It learned only from grown-up voices.
So it works well for grown-ups. It often fails for kids.
That is not fair. The data had a gap: kid voices.
Real AI can have this problem too. One-sided data can leave some people less well served.
A lean toward some people that is not fair is called bias.
| Voices in the training set | Number of examples |
|---|---|
| Grown-up voices | 18 |
| Kid voices | 0 |
Training data is usually picked by people.
So people can make it fair, or they can miss a gap by accident.
A good habit is to ask: who or what might be missing?
| Statement | True or false? |
|---|---|
| One-sided data can make AI work worse for some people. | ? |
| People usually choose the training data. | ? |
| If a tool fails for you, it must be your fault. | ? |
| Asking who is missing helps find a gap. | ? |
Thoughtful work, explorer! Tomorrow you build a one-sided model on purpose, then fix it.