Designing AI Features That Users Can Trust, Not Just Use
Most advice on AI features focuses on making them feel smart. I think that's the wrong goal. A feature that feels smart but can't be checked will be trusted too much by some people and not at all by others. Both are bad outcomes.
I'm a product designer, not an ML engineer, so I won't pretend to know how to make a model more accurate. What I can reason about is the interface around it. That's where I think trust is won or lost.
01Trust is about calibration, not confidence
The goal isn't maximum trust. It's calibrated trust: users rely on the feature when it's reliable and double-check when it isn't.
Say a tool summarizes a long support thread. If the summary looks polished and authoritative, users will assume it's complete. Then one day it misses a key detail and they stop using it entirely. If the interface had made it easy to jump from each summary line back to the original message, they'd have caught the miss and kept trusting the rest.
So the question I'd ask for every AI feature is: how will a user know when this is wrong?
02Show where the answer came from
The strongest trust cue I know of is a visible source. If the output is based on the user's own data, such as documents, messages or records, link back to it.
- Put the reference next to the claim, not in a list at the bottom.
- Make the link open the exact place, not just the parent document.
- If there is no source, say so. "Generated from general knowledge" is more honest than silence.
This costs real engineering effort, and I can't tell you how hard it is for your stack. But as a design requirement, I'd push for it early, because it's very hard to bolt on later.
03Be specific about limits, near the action
"AI can make mistakes" in a footer is a legal sentence, not a design decision. Nobody reads it, and it doesn't tell anyone what to do.
Better is a short, specific note at the moment it matters:
- Before sending: "Review names and dates before you send this."
- On a generated chart: "Based on the last 30 days of data only."
- On a draft: "This is a first draft. It hasn't been fact-checked."
Specific cues change behavior because they point to what to check. Generic ones are just noise.
I'd also avoid sprinkling warnings everywhere. If every output carries a caution, users learn to ignore all of them. Save them for outputs with consequences.
04Confidence scores: handle with care
It's tempting to show a percentage. It looks scientific. But I'd be cautious, for two reasons.
First, users read a number as a guarantee. "92% confident" feels like it will be right nearly every time, and that's rarely a claim a product can support.
Second, a number doesn't tell people what to do. What should they do at 74%? Nobody knows.
What I'd try instead is flagging the specific parts that need attention, for example highlighting a field the tool had to guess, or marking an extracted value as "unclear in source." That turns uncertainty into an action.
05Make everything reversible
Trust is much easier to give when mistakes are cheap. This is old UX wisdom, and it matters even more with AI, because output can be wrong in ways users don't notice immediately.
A few patterns I'd reach for:
- Preview before apply. Show what will change, then let the user confirm.
- Undo that really works. Not a toast that disappears in three seconds. If the AI edited twenty records, undo should cover all twenty.
- Keep the original. When AI rewrites something, let users compare or restore the previous version.
- Separate suggestion from action. A draft the user edits is lower risk than something sent automatically.
The more an action affects other people or can't be taken back, the more friction I'd add before it happens. Sending an email to customers deserves a review step. Rewording a heading doesn't.
06Let users correct it, and show that it mattered
If a user fixes an output, the interface should respond. A thumbs-down button that does nothing visible feels like shouting into a void.
I can't tell you whether your system actually learns from corrections, and that's the point: don't imply it does if it doesn't. If edits only change the current result, say that. If they genuinely improve future results, say that too. Promising learning you don't deliver is a quick way to lose people.
Even simple things help: let users edit the output inline, regenerate with a nudge like "shorter" or "more formal," and keep their edits when they regenerate. Losing a user's manual changes because the AI re-ran is a small betrayal that people remember.
Also plan for the failure moments. I covered how in Designing Loading and Error States for AI Features. How a feature behaves when it breaks says more about trust than how it behaves when it works.
07A short checklist before you ship
Run through these questions for each AI feature:
- How would a user discover that the output is wrong?
- Can they trace any claim back to a source?
- Are limits explained at the point of use, in specific words?
- Is every consequential action previewed or reversible?
- Do user corrections stick, and is it honest about what happens next?
- Does the feature stay quiet when it has nothing reliable to say?
That last one is underrated. Sometimes the most trustworthy thing an AI feature can do is say "I don't have enough information to answer that," instead of producing something plausible. Design a good-looking version of that response. It's a feature, not a failure.
08My take
I think teams overinvest in making AI output impressive and underinvest in making it checkable. Users forgive a tool that's sometimes wrong and tells them how to verify it. They don't forgive one that is confidently wrong and hard to question.
If you only have time for two things, do sources and undo. Everything else builds on those.
09Frequently asked questions
Should an AI feature show a confidence score?
Usually not as a bare percentage. People read "87%" as a promise, and most products can't back it up. Showing sources, or marking which parts need checking, gives users something they can act on.
How do I make users trust an AI feature without overselling it?
State what it can and can't do near the point of use, show where answers come from, and make every action reversible. Trust grows when the feature is honest about its limits.
Is a disclaimer like "AI can make mistakes" enough?
No. A generic footer line is easy to ignore and gives users no guidance. Put specific cues next to the output that matters, such as "Check the dates before sending."
Denny Pratama
Product designer in Indonesia. Since 2018 I’ve designed SaaS products, dashboards and marketing sites, and now I build my own designs too.