Rocket sketches the first Sky Sorter design on the whiteboard. A big arrow points to one box marked ANSWER.
Raven taps the box with her marker. "Answer to what, exactly? It could say a lot of different things."
"It could guess what the card shows," Rocket offers. "Or suggest which cards a scientist should look at first."
"Or it could just drop each card into a bin," Raven adds.
Nova hovers beside the board. "Those are three different jobs," she says. "Your design should name which one Sky Sorter does."
Rocket erases ANSWER and writes three new words underneath.
The output of an AI system is what it gives back. The definition names three common kinds.
A prediction says what something probably is, or what will probably happen.
A recommendation suggests what someone might pick, try or do next.
A decision settles what actually happens, such as which bin a card goes into.
One system can produce more than one kind, so designers must say which outputs they want.
| Kind of output | Sky Sorter example | Signal words |
|---|---|---|
| prediction | This card is probably a comet. | probably, likely, will |
| recommendation | A scientist might want to check this card first. | might, try, suggest |
| decision | The card is filed in the comet bin. | filed, sent, approved |
Every AI system is built for a set of objectives. An objective is a goal that people choose.
Mission control's objective is clear: label every snapshot card correctly.
A different team might set a different objective, such as finding all comet cards as fast as possible.
The objective shapes which outputs matter. It also shapes how people will judge whether the system works.
| Statement | True or false? |
|---|---|
| An objective is a goal chosen by people. | ? |
| A recommendation always makes the final choice for you. | ? |
| One system can produce more than one kind of output. | ? |
| A prediction can be wrong. | ? |
Nice work. Tomorrow's Explorer Lab turns a dial on how much Sky Sorter does by itself.