Designing for answers that might be wrong

Most of the screens I designed before 2023 had a quiet assumption underneath them: if the system shows something, it’s true. A balance is a balance. An order status is an order status. The design work was about clarity, not credibility.

AI features break that assumption. Ask the same question twice and you can get two different answers, both sounding equally sure of themselves. So I’ve had to change what I think a good screen is for.

Four patterns I now reach for first

Show it as it arrives, but don’t fake it. Showing an answer word by word as it’s written makes a slow answer feel faster, and that’s worth having. What I avoid is a fake typing effect on text that’s already finished. It reads as theatre, and people stop reading and start waiting for it to end. Show it as it arrives when it’s genuinely arriving, and all at once when it isn’t.

Put sources next to the claim, not under the answer. A list of links at the bottom is decoration. Nobody clicks through four of them. On a support tool I worked on, the panel only started to feel trustworthy once hovering over a sentence showed the exact line it came from. If there’s nothing to point at, the AI should say so, and the screen should treat that as a normal state, not an error.

Use confidence to tell people where to check, not to give a score. Raw percentages invite false precision: is 71% good? I’ve had more luck with a few plain labels (“checked against policy”, “no source found”, “needs a human”), each with a visible reason. The point is to tell people where to look, not to grade the AI.

Put “try again” next to “edit”. A “try again” button on its own is a slot machine. Pair it with direct editing and keep the previous version one tap away, so people can compare rather than gamble.

Undo is the real trust feature

The biggest change in how I work: I design the undo before I design the action. If an assistant can send, book, spend or delete, I want to know three things first. What exactly gets reversed? For how long? And what does the “here’s what I did” receipt look like?

People will let a system act on their behalf if they believe they can take it back.

In my experience, a real undo window does more for trust than any amount of careful confirmation copy. If people can’t take it back, they’ll second-guess every step.

There’s a cost to this. Undo has to be real. Engineers have to build a genuine way to reverse things, rather than “we’ll email support”. That’s a conversation I now have in week one, not at handover.

The takeaway

Don’t design an AI feature as if it’s always right. Design for a sensible person who expects it to be wrong sometimes. Make checking quick, make mistakes cheap, and let people step in before the system commits.