Teacher Lin stops by with a stack of Biology homework. "If Check marks these," she says, "maybe I can stop reading answers altogether."
Comet's eyes light up. "A teacher plus SH-1. That has to beat either one alone!"
Wren frowns at his clipboard. "What does the evidence say? On our test, Check missed 3 of 10 wrong answers. If nobody reads, those 3 slip through."
"So together could be worse," Comet says slowly, "if we team up the wrong way."
Nova draws two overlapping circles, then shifts them apart and back. "Would you like a hint?" she asks. "Ask how the team is organized, not just who is on it."
Teacher Lin smiles at you. "Reviewer, how should we team up?"
Working together, a person and an AI can sometimes be more biased than either one alone.
Organized well, they can do better than either one alone.
How much people are empowered to challenge AI output still needs study.
Generative AI may call for extra human review, tracking and oversight. Organizations can restrict AI uses that cause harm.
| Made-up team set-up | What happens to Check's 3 missed errors (CK-20) |
|---|---|
| Teacher stops reading; Check alone marks homework | All 3 get through. Students keep 3 wrong ideas. |
| Teacher reads every answer; Check adds a second opinion | The teacher can catch them. Check may still help spot errors the teacher misses. |
Transparency answers what happened. Explainability answers how a decision was made.
Interpretability answers why, and what it means to the user.
Explanations can be fitted to the user's role, knowledge and skill. A student, a teacher and the panel may each need a different one.
| Statement | True or false? |
|---|---|
| A person and an AI together are always better than either alone. | ? |
| Organized well, a person and an AI can do better than either alone. | ? |
| Organizations can restrict AI uses that cause harm. | ? |
| Every user needs exactly the same explanation. | ? |
| Transparency answers the question of what happened. | ? |
Sharp thinking, reviewer. Tomorrow is Impact Friday: who is accountable when SH-1 gets it wrong?