Designing AI interfaces people trust
Most of the AI interfaces I get asked to review have the same shape. A text field, a button, then a clean block of copy. Designing AI interfaces like that treats every answer as a fact. Plenty of them are guesses in a nice typeface.
An interface people can trust shows three things before it asks anyone to act: where the answer came from, how sure the system is, and what to do when it is wrong. The model can stay in the background. The screen has to do this work.
I have designed products in this space, not only homepages about them. K_AI helps visitors find art at Staatsgalerie Stuttgart. For SYNIO, a Stuttgart startup that trains models on synthetic images, I did the site and the brand. In both projects the awkward questions showed up after the demo already looked good.
Where designing AI interfaces slips
Tools that turn a prompt into a whole screen are fast. They are also why so many new products look related. Same card grid. Same soft shadow. Same hero, gradient, one button. The screen runs. It does not belong to anyone.
People notice, even when they cannot name it. A bank, a museum, and a training tool should not share a layout because the same generator touched them. If you already have a design system, the generated screen has to sit inside it. Type, spacing, and components are part of whether the product feels deliberate. An off-brand color is a small crack until the user starts wondering what else was improvised.
I still sketch the flow before anyone prompts a screen into existence. The prompt helps once the job is clear: what the person came to do, what they are afraid of getting wrong, and which step is allowed to be slow.
Confidence is a design choice
A wrong answer in a plain system font does less damage than a wrong answer that looks like a finished report. Fluent sentences read as certainty. If the solid answer and the shaky one use the same bubble, the same weight, and no source, people treat them as the same kind of thing.
In a museum tool that split matters. Suggesting a painting someone might like is a soft recommendation. Stating when it was made, or who sat for it, is a claim. Those two outputs should look different. One can be a short list with a reason. The other needs a source a visitor can check, or it should stay unsaid.
When I review a flow I sort each AI output into one of three piles:
- Retrieved: the system looked something up and can point at it.
- Suggested: the system is offering a next step, and ignoring it is fine.
- Unsure: the screen should say so, in the interface, not in a footnote three scrolls down.
A percentage badge on every line does not help. Most people have no use for “73% confident.” They do know what to do with “this is in the collection catalogue” and “this is a guess, here is someone who can check it.”
What I put on the screen
A source, or an honest gap
If the answer leans on a page, a record, or a document, link it. If there is nothing to link, say that the system is inferring. A smooth paragraph over that gap is how a product earns a calm, wrong reputation.
A next step other than regenerate
Regenerate is a shrug. Narrower moves work better: change one part, ask a person, show the three closest matches, or fall back to ordinary search. On K_AI the useful job was helping someone wander the collection. A chat that had to be right would have fought that.
The failure, drawn first
I draw the empty state, the slow state, and the wrong state before the happy path. Those are the screens people remember. A spinner with no sense of time feels broken. A wrong answer with no way back feels like a lie. Both need a visible exit.
Fewer choices when the stakes rise
For a low-stakes suggestion, a few options side by side are enough. When someone might repeat the output to a visitor, a customer, or a colleague as fact, the interface should slow them down. Ask for a check. Show the source again. Speed is welcome until it helps a mistake travel.
Start from the job, then the model
Teams often show up with a model already picked and a blank page where the product should be. I turn that around. We write the task in one sentence, the moment the system may be wrong, and the moment it must not be. Then we decide what the screen shows.
That is also how I scope UI and product work when AI is part of it. People judge the interface. The model is a supplier.
If you are adding an AI feature and the current version is a prompt box with a perfect-looking answer, write to me. I am in Stuttgart and I work with teams on the design and the build. The contact page is the short route.
Comments are closed.