Designing AI features people can trust
By Imran Sheikh · 2026-11-12 · AI in practice · 4 min read
Key points
- Show sources so people can check the answer themselves.
- Say when the system is unsure, and say it plainly.
- A good refusal still helps the person move forward.
People do not trust an AI feature because the model is clever. They trust it because the product has been honest with them, again and again, in small ways. Trust is a design problem, and a solvable one.
Three patterns carry most of it: citations, uncertainty and refusals. Add a fourth, which is control. Together they tell a person what the system knows, what it does not, and what it will not do.
Citations give people a way to check
An answer without a source asks for blind faith. Some people give it. Most do not, and the ones who do get burned eventually. A source next to a claim turns faith into something the person can verify in a few seconds.
Good citations are close to the claim they support. A tiny numbered marker after the sentence works well. Tapping it should show the passage, the document name and a way to open the original. A list of links at the bottom of a long answer is much weaker.
- Attach each source to the specific claim, not to the whole answer.
- Show the actual passage, so people can see what was used.
- Link to the original document when one exists.
- Say plainly when an answer has no source behind it.
- Never invent a citation. A missing source beats a fake one.
That last point is a design rule as much as an engineering one. If the system cannot ground a claim, the interface should show that. A visible gap builds more trust than a confident answer that falls apart on inspection.
Uncertainty should be visible and plain
Models sound sure even when they are guessing. The interface has to correct for that. If the system is unsure, the person needs to know, and in words they understand.
Avoid false precision. A percentage confidence score looks scientific and usually means little. Plain language does better. This looks likely. I could not find a clear answer. These sources disagree.
Show what would settle the question. If the answer depends on a detail the person has not given, ask for it. If it depends on something the system cannot know, say so. Uncertainty paired with a next step feels helpful. Uncertainty alone feels like a shrug.
Design the refusal
Every AI feature will refuse some requests. Some are out of scope. Some are unsafe. Some the system simply cannot do well. Most products treat the refusal as an error message. It is a moment of design, and it shapes how people feel about the whole product.
A good refusal is short, kind and specific. It says what the system will not do. It gives a reason in plain words when it can. Then it offers something useful instead, such as a related answer, a different path or a human to talk to.
- Say what you cannot do in the first sentence.
- Give a brief, honest reason without lecturing.
- Offer the nearest thing you can help with.
- Point to a person or a resource when the stakes are high.
- Keep the tone the same as the rest of the product.
A cold, generic refusal makes people feel accused. A warm, specific one makes them feel looked after. The facts are the same. The experience is not.
Keep people in control
Trust also depends on power. People trust features they can steer, undo and correct. Let them edit the output. Let them reject a suggestion in one tap. Let them see what the system used to reach an answer, and change it.
Show what the system remembers about them, and let them delete it. Say what leaves the device and what stays. Plain privacy language placed at the moment of need works better than a long policy no one reads.
Control also means pace. Let people stop a long answer midway. Let them ask the system to try again, to be shorter or to explain a term. Each of these gives the person a hand on the wheel, and a person with a hand on the wheel is far more willing to keep driving.
A note on health and companion products
These principles matter most when the subject is personal. We make Emerge Care, an AI health companion, and the design questions there are about tone and honesty more than features. The product should sound warm and steady. It should be clear about what it is and what it is not.
- State plainly that the companion is an AI and not a clinician.
- Encourage people to talk to a qualified professional about medical decisions.
- Ground factual statements in sources people can read.
- Make a path to real help easy to find, especially in difficult moments.
- Let people control and delete what the app knows about them.
These are general design principles and make no medical claims. The aim is an interface that is honest about its limits. People in vulnerable moments deserve exactly that.
Test trust directly
You can test these choices. Put a build in front of people and ask them what they would do with the answer. Would they act on it. Would they check it. Where do they hesitate. Watch whether they open the sources, and whether they notice the uncertainty.
Review the failures as carefully as the wins. When the system is wrong, what does the person see. Can they tell. Can they recover. A product earns trust in those moments, when something goes wrong and it handles it with honesty.
Trust is built from small admissions, made consistently.