AI and user experience: designing for calibrated trust
AI changes user experience because the interface no longer gives the same answer to the same question: it gives probable answers. The design goal is no longer maximum trust but calibrated trust: users who know when to act, when to check and when the AI simply doesn’t know.
What changes when an interface stops being deterministic?
For decades, interface design has rested on a simple promise: the same button always does the same thing. A filter returns the same rows; a form validates the same way every time. You learn the tool once and you trust it.
Put a model inside and that promise breaks. The same input can produce a different answer, and a wrong answer sounds exactly like a right one: same tone, same layout, same confidence.
That asks something new of your users: knowing how far to trust what they see. If the interface doesn’t tell them, they’ll work it out for themselves, and usually get it wrong.
The design side of our team comes from creative and UX direction at a national telecoms operator and from projects for major banking, energy and financial services firms. In those settings, a confusing screen already cost calls and mistakes. With AI, a screen that looks surer than it is can cost a decision.
What is calibrated trust, and why is it the goal?
Calibrated trust means users trust each answer as much as that answer deserves. No more, no less. It fails in two ways:
- Overtrust: users accept everything, mistakes included. The AI is effectively deciding, even though nobody decided it should.
- Undertrust: users check everything or ignore everything, correct answers included. The AI adds nothing and still costs money.
The second is more common than it looks. An alert nobody can check ends up muted, and an assistant that was confidently wrong twice stops being used, even if it gets the next nine right.
So the goal isn’t “get people to trust the AI”. It’s getting them to trust it when that trust is warranted. That is the visible face of what we call accountable AI: every answer comes with its evidence, or says it isn’t sure.
And you can measure it. In a pilot, watch two numbers: how many correct answers your team discards and how many wrong ones it accepts without checking. If both go down, the interface is calibrating well.
Which design patterns work with AI?
There’s no magic component. There’s a handful of patterns that, together, give users what they need to decide.
Show how far to trust it
Every answer carries its level of confidence, in language the user understands: high, medium, low, or a range. Beware of percentages with decimals: they suggest a precision the model often doesn’t have. And if the uncertainty isn’t calibrated, say so.
Show the evidence
The answer comes with what supports it: the citation to the document and section, or the variables that weighed most in the prediction. Evidence isn’t decoration. It’s what lets someone check in seconds instead of redoing the work.
Let the AI abstain
“I don’t know” is a feature, not a failure. A system that warns when the data is outside what it knows, or that finds no support and says so, spares users the worst experience of all: acting on a made-up answer.
Make correcting and overriding easy
If users know the AI can be wrong, they need to be able to correct it without fighting the interface. Overriding a decision should take one action, be logged, and not reappear on the next screen as if nothing had happened.
Explain in layers
First the decision. Then how far to trust it. Then why. Finally, the raw data. Someone working under pressure sees the first layer; someone auditing gets down to the last one without switching tools.
Don’t humanise it or fake confidence
No “let me think”, no effusive apologies, no tone that sounds just as certain about the shaky as about the solid. An interface that imitates a person invites the kind of trust you give a person, and that trust isn’t calibrated.
Respect the latency budget
Not everything fits in the same time frame. A verdict on a reading has to arrive in milliseconds, on the data path. A language model takes seconds: its place is on top, explaining what happened, not in the middle of every reading.
Let a person approve what’s sensitive
The AI proposes; a person approves anything with consequences. That’s what human in the loop means in practice, and the design has to make clear what has been proposed and what has been done.
Make it clear without a manual
If using the AI well takes a training course, the interface is pushing its complexity onto the user. Confidence, evidence and abstention have to be readable at a glance.
What does each pattern look like in practice?
| Pattern | What the user sees | Why it matters |
|---|---|---|
| How far to trust it | High, medium or low confidence, or a range | Separates what they can use from what they should check |
| Evidence | A citation with document and section, or the variables that weighed in | They check in seconds instead of redoing the work |
| Abstention | “No supporting source” or “data outside what the model knows” | Nobody acts on a made-up answer |
| Correction and override | One control to correct, with a log | The person stays in charge and the mistake doesn’t repeat silently |
| Layered explanation | The decision first, detail on demand | Works for the operator in a hurry and for the auditor |
| No humanising | A neutral tone, equally clear about the certain and the doubtful | Doesn’t invite trust that hasn’t been earned |
| Latency | An instant verdict, the explanation afterwards | Each thing arrives when it’s useful |
| Human approval | Proposals labelled as proposals | Nothing sensitive happens without someone deciding it |
How do our products do it?
For us these patterns aren’t theory: they’re product decisions.
In CogniX, our predictive AI, every answer comes in four parts: the decision, how far to trust it, a warning if the data is outside what it knows, and which variables weighed in. Uncertainty is calibrated where the maths allows, and it says where it doesn’t. The explanation is exact where the model allows and marked as indicative where it doesn’t. None of it is sold as a cause.
In RAGNIX, document retrieval with verifiable citations, the rule is “cite or abstain”. Every answer carries the document, revision and section; it never hands over half a procedure, and if it finds no support, it says so. If something fails, it says that too, rather than degrading silently.
In NovaeOps, our AIoT platform, CogniX’s verdict travels inside the readings you choose, with the same four parts. If the AI is unavailable, the reading is still stored: the AI adds something, but doesn’t break what already worked. And the verdict informs; you decide, with rules, what triggers an automatic action.
Where do you start?
Pick the screen where your team already uses AI, or is about to, and ask three questions. Can you see how far to trust it? Can you see where it comes from? Can the AI say it doesn’t know? If any answer is no, that’s your first change.
If you’d like to do it with a team that designs the experience and builds the system, see how we approach applied AI, from pilot to production.