Mission control has called a special meeting. A sign on the door says "AI Review Board: Sky Sorter".
Rocket straightens his collar. "Are we in trouble?"
Raven sets a folder on the table. "No. Before any system is used for real, someone should check it carefully. That someone is us."
Nova projects Sky Sorter's whole story onto the wall: features, training cards, test scores and the bias the crew found.
"Today we decide what trustworthy means," she says. "Tomorrow we start checking Sky Sorter against it."
Rocket picks up a pencil. "Then I want a really good checklist."
A US standards agency, NIST, lists the characteristics of trustworthy AI.
Some come in pairs. Here they are, each with a question the review board asks about it.
| Characteristic | A question the board asks |
|---|---|
| Valid and reliable | Does it do its job right, again and again? |
| Safe | Could it hurt anyone? |
| Secure and resilient | Could someone misuse it, and can it recover if something goes wrong? |
| Accountable and transparent | Can people find information about it, and who is responsible for it? |
| Explainable and interpretable | Can people see how it works and what its outputs mean? |
| Privacy-enhanced | Does it protect people's personal information? |
| Fair, with harmful bias managed | Has anyone looked for bias and worked to reduce harm? |
| Statement | True or false? |
|---|---|
| Trustworthy AI should have harmful bias managed. | ? |
| If an AI system is accurate, nothing else matters. | ? |
| Someone should be responsible for an AI system. | ? |
| A review board only checks whether a system is fast. | ? |
The board is ready. Tomorrow you learn what it means to see inside a system.