After the camera change, the crew retrained Sky Sorter on a new box of forty labeled cards.
Rocket pins the results of a bigger test, twenty fresh cards, to the mission board. "15 out of 20 right! That is a great score."
Raven runs her finger down the results table. Then she stops at the bottom rows.
"Look at the comets," she says quietly. "Two comet cards in the test, and Sky Sorter missed both."
Rocket frowns. "But nobody told it to ignore comets. We never meant to be unfair."
Nova hovers beside the board, her lights pulsing slowly. "Meaning to be unfair and being unfair are different things," she says.
"What if the problem started before Sky Sorter ever saw a test card?"
In everyday talk, bias means leaning toward one side. In AI, bias is a steady lean that makes results unfair or wrong for some cases.
Bias is not the same as a random mistake. A random mistake could land anywhere. Bias keeps landing in the same place.
Sky Sorter missed both comets. That is a pattern, not bad luck, so it is worth a closer look.
NIST, the US National Institute of Standards and Technology, wrote a framework for managing the risks of AI.
It names three major kinds of AI bias: systemic, computational and statistical, and human-cognitive.
It also says each kind can happen without anyone meaning to be unfair.
This week you will meet all three and find each one hiding in Sky Sorter.
| Statement | True or false? |
|---|---|
| NIST names three major kinds of AI bias. | ? |
| Bias only happens when someone wants to be unfair. | ? |
| A high overall score proves a model works well for every kind of case. | ? |
| Human-cognitive bias is about how people read and use an AI system's output. | ? |
Strong start, explorer. Tomorrow you open the training box and count what is inside.