AI in practice · Mixed

Designing an interface for a system that isn't always right

In one line: A good interface shows varying certainty, allows easy correction, and doesn't present a guess as a fact.

Show the source

An answer with a link to the passage it was drawn from lets the user verify in two seconds. That changes the relationship: from “trust me” to “here, check.”

When there's no confidence, say so in the interface — and don't mimic certainty in the design.

Correction as part of the flow

Allow editing the output in place, rejecting it, or asking for a rephrasing. Every correction is also a signal you can collect.

For actions with impact — show, confirm, execute. Not automatic on one click.

Expectations

Briefly explain what the system knows and doesn't, and show progress during long processing. Wording matters too: “suggestion” invites scrutiny, “the answer” invites blind trust.

Going deeper

Explicitly design the failure state: no answer, no confidence, service unavailable. In model-based systems these aren't rare edge cases but a constant part of use, and an interface that ignores them loses trust the first time they occur.

Users

Next: Bias, fairness and accountability in using models · Previous: Data: where it comes from and what to do when there is none · Back to library