BlogProduct design

Product design

Showing Uncertainty in an AI Interface Without Confusing Users

Give users practical signals about missing information and unresolved choices so they can decide what to verify, correct, or hand off in an AI workflow.

Users do not need a warning attached to every generated sentence. They need to know when an unresolved issue changes what they should do next. Design uncertainty around specific gaps in the task: missing information, conflicting sources, an ambiguous request, or an action that needs someone else’s judgment.

Explain the reason, not just a vague label

A label such as “low confidence” leaves the user to guess what is wrong. A message such as “the product revision was not included” points to a fix. Prefer explanations tied to observable conditions in the workflow rather than unexplained numerical scores.

In a hypothetical service assistant, two guides might describe different installation procedures for different revisions. Show that distinction and ask the user to choose the relevant product. Do not hide the conflict inside a fluent combined answer that makes the information seem more settled than it is.

Match the interruption to the decision

Keep minor issues close to the affected field or passage. Reserve stronger interruptions for moments where continuing could create a consequential mistake. The interface should help the user focus on the particular issue rather than making every result look equally problematic.

Provide an action alongside the explanation: add a missing detail, inspect the original document, choose between alternatives, or ask a colleague. If the product cannot proceed responsibly, preserve the user’s work and explain the next available route. An unhelpful refusal can be as frustrating as an unsupported answer.

Test whether people interpret the signal correctly

Show realistic examples to users and ask what they would do next. Check whether they distinguish an incomplete input from a problem with the source information. A design that seems clear to its authors may suggest something different to the people relying on it.

Review signals after launch alongside corrections and support requests. If users routinely ignore a notice, investigate whether it appears too often, explains too little, or interrupts the wrong moment. Improve the signal based on the decision it should support instead of adding more alarming language.