Designing trust into AI products
The difference between AI people use and AI they abandon is rarely the model. It's the experience around it.
The most accurate model in the world is worthless if people don't trust it enough to act on it. Trust is a design problem as much as an engineering one.
Show your work
Citations, confidence and explanations turn a black box into a colleague. When an AI shows why it reached a conclusion, users can verify it — and verification builds trust fast.
Two ways to be uncertain
Low confidence
- Loses trust
- Answers anyway, in the same tone
- Keeps it
- Says what it doesn't know, and why
Sources disagree
- Loses trust
- Picks one silently
- Keeps it
- Surfaces the disagreement
No supporting evidence
- Loses trust
- Improvises plausibly
- Keeps it
- Declines and offers a human
It was wrong
- Loses trust
- No way to tell how it got there
- Keeps it
- Citations you can check
| Loses trust | Keeps it | |
|---|---|---|
| Low confidence | Answers anyway, in the same tone | Says what it doesn't know, and why |
| Sources disagree | Picks one silently | Surfaces the disagreement |
| No supporting evidence | Improvises plausibly | Declines and offers a human |
| It was wrong | No way to tell how it got there | Citations you can check |
Design for the wrong answer
AI will be wrong sometimes. Great AI products make it easy to catch and correct mistakes, so a wrong answer is a small friction, not a disaster.
Design for the wrong answer
- 1
Assume it will be wrong sometimes
Because it will. The question is what the user can do about it.
- 2
Make checking cheap
Citations that open the source in one click.
- 3
Make correcting easy
Edit and resend beats starting a new conversation.
- 4
Capture the correction
It is the most valuable signal the product produces.