AI × UX·3 min read·

Designing for the moment of doubt

Most AI products are designed for the moment the machine is right. The design that matters is for the moment a person has to decide whether to believe it.

An AI answer arrives looking finished. A clean sentence, a confident tone. Nothing on the screen tells you whether it came from two hundred rows of data or from a guess.

And someone is about to act on it. Forward it, put it in a report, make a decision with it.

That gap is the design problem. Not the chat box. The moment when someone has to decide what to do with what they were told.

Doubt does not wait for the answer, though. It shows up at four points, and each one can be designed.

Before the question

Doubt starts before anyone types.

Picture two assistants. The first is an empty box that says “Ask me anything”. The second says “I read your spreadsheet and answer questions about it”.

With the first, you have no idea what to expect, so you cannot tell a good answer from a bad one. With the second, you know its job, and you can tell at once whether it did it.

So every assistant I design gets that one plain line. It does more than the usual warning. A warning tells people the product is unreliable. A statement of scope tells them how to use it.

While it works

Some tasks take a minute. A minute in front of an unchanging screen is long enough to start wondering whether anything is happening at all.

So on a long task the screen stays alive: what the assistant is doing now, and for how long. “Building your slides · 0:12” is enough. People who do not want the detail can fold it into a single line, but the line keeps moving.

When the answer lands

The same answer can be shown two ways. “Supplier C caused the most delays” is a sentence you either trust or you don’t. Add where it came from, how sure the system is, and what it could not see, and it becomes something you can weigh.

I would rather an answer say plainly “this does not include the last three weeks” than sound sure about something it has not seen. Admitting the gap is what makes the rest believable.

When something goes wrong

The worst design here is one red message for everything. I separate three cases.

  • Something broke: say so, and offer to try again.
  • The answer is incomplete, but nothing failed: say what is missing.
  • The request cannot be done as asked: say what can be done instead, and never offer “Try again”, because a button that cannot work costs trust.

Even stopping has two meanings. If you watched the answer stop, you already know what happened, so the screen keeps what was written and offers to continue. If you left and came back to nothing, you need to be told why.

What it adds up to

None of this makes the model smarter. It gives the person in front of it what they need to decide.

A model that is right most of the time still needs a design for the times it is not.

The interface is easy. The moment of doubt is the design.

Samer HaddadUI/UX Manager, leading design for AI products