Designing for uncertainty: AI that's honest about what it doesn't know

AI features don't return facts, they return probabilities dressed as facts. Designing for uncertainty means showing how sure the system is, where its answer came from, and how to recover when it's wrong — so a confident-sounding mistake doesn't quietly become the user's problem.

Why is a confident wrong answer so dangerous?

People trust fluent language. An AI that states a wrong answer in the same calm tone as a right one borrows credibility it hasn't earned. The damage isn't the error; it's that the user had no signal to doubt it.

How do you make uncertainty visible without clutter?

Not with a percentage on everything. Show confidence where it changes what the user should do next.

• Cite sources the user can open and check.
• Hedge the copy when the system is unsure, and commit when it isn't.
• Offer a second option rather than one false certainty.
• Make correction fast — edit, regenerate or undo in one step.

What happens when it's wrong?

Recovery is a design surface, not an error state. A good AI feature assumes it will sometimes be wrong and makes being wrong cheap: easy to spot, easy to fix, never destructive by default.

How we use it

We treat uncertainty as content to design, not a flaw to hide — because a feature people can trust when it's unsure is one they'll keep using.