AI for UX · Step 5 · 5 min read
Designing for uncertainty
AI is sometimes wrong, and people need to know when. How to show confidence without fake precision, why explanations can backfire, how to word uncertainty, how to design errors and graceful failure, and how to help people trust AI exactly as much as it deserves.
Traditional software is either right or broken. AI is usually right, sometimes wrong, and always sounds sure. That's a new design problem: people need to know how far to trust each answer, how to check it, and what to do when it's wrong. Too little trust and nobody uses the feature; too much and people stop checking. The goal is in between: trust that matches reality.
What you'll be able to do after reading:
- Explain trust calibration and why both over-trust and under-trust are failures
- Choose a way to show confidence: categories, alternatives or numbers
- Word uncertainty so people actually check
- Design errors and graceful failure for AI features
- Know when explanations help, and when they make things worse
1. The goal: calibrated trust
Google's PAIR guidebook puts it plainly: help users calibrate their trust; "the user shouldn't trust the system completely" (PAIR). Microsoft's guidelines start the same way: "Make clear how well the system can do what it can do" (Amershi et al., 2019).
People ignore good suggestions and redo the work.
The feature is wastedPeople rely on the AI where it is good and check where it is weak.
The goalPeople stop checking and accept errors.
Automation bias
See Trust calibration.
2. Showing confidence
PAIR describes three ways to display how sure a system is (PAIR):
High · medium · low, each saying what to do next.
Usually the clearestThe top few options instead of one answer.
Best when confidence is lowA percentage or score.
Hard to understand; test itPAIR also warns that confidence scores can be hard for people to understand, so test the display early. In practice, categories that say what to do next usually beat raw percentages. See Confidence indicators.
3. Words matter: first-person uncertainty
How the AI expresses doubt changes what people do. In a pre-registered study with 404 participants, answers that said "I'm not sure, but…" made people agree with the AI less often and made their final answers more accurate, compared with impersonal wording like "It's not clear" (Kim, Liao & Vorvoreanu, 2024). NN/g recommends the same for hallucination-prone features (NN/g).
Practical wording:
- "I'm not sure about the tax code on this invoice. Please check line 4."
- "I found two possible matches. Which one did you mean?"
- Not: "Tax code: 0471" in the same style as everything else.
4. Explanations: helpful, and risky
Explanations ("why did it suggest this?") seem like an obvious good. The evidence is mixed:
- In one study, explanations increased how often people accepted the AI's recommendation, whether it was right or not (Bansal et al., CHI 2021). A convincing reason can make a wrong answer more persuasive.
- NN/g warns that the explanations AI chatbots give about their own reasoning are often inaccurate (NN/g).
So: prefer explanations that point to checkable evidence (the source document, the data used, the rule applied) over explanations that sound reasonable. Microsoft's guideline: "Make clear why the system did what it did", in a way people can verify. See Explainability.
5. Errors and graceful failure
PAIR asks teams to decide what counts as an error at all: a recommender that is useful 60% of the time can be a failure or a success, depending on the task (PAIR). Then design for it:
| Guideline | In the interface |
|---|---|
| Support efficient correction (Microsoft G9) | Edit in place, undo, "not this one" next to every AI suggestion |
| Scope services when in doubt (G10) | Ask a clarifying question, or offer fewer, safer options, instead of guessing |
| Provide a path forward (PAIR) | Every failure ends in something the person can do: retry, choose manually, contact a person |
| Explain the limit | "I can't read handwritten invoices yet. Upload a typed copy or enter it manually." |
See AI error states.
6. Vocabulary
| Term | Meaning |
|---|---|
| Trust calibration | Trust that matches how reliable the system really is |
| Over-trust / under-trust | Relying on AI more, or less, than it deserves |
| Confidence indicators | Ways of showing how sure the AI is |
| N-best alternatives | Showing the top few options instead of one answer |
| Explainability | Making clear why the AI produced a result |
| Graceful failure | Failing in a way that still leaves the person a good next step |
7. Try it (no code needed)
- Pick one AI feature you use or design.
- Find a case where it's likely to be unsure (a blurry scan, an ambiguous request).
- Design three versions of that moment: a confidence category, two alternatives, and a first-person "I'm not sure" message.
- Ask two colleagues which one makes them check the result. Note why.
Reflection questions: Where does your product currently present an uncertain answer as certain? What would "provide a path forward" look like on your most common AI error?
8. If you lead a team
- Add confidence and correction to the AI pattern library: badges, alternatives, undo.
- Require every AI feature to define what counts as an error, and the path forward for each.
- Test explanations with deliberately wrong AI outputs, to see whether they make people accept errors.
Checklist
Learn more
- Guidelines: Google PAIR: Explainability + Trust · PAIR: Errors + Graceful Failure · Microsoft HAX Toolkit
- Research: Kim et al., "I'm not sure, but…" (FAccT 2024) · Zhang et al., confidence and explanation (FAT* 2020) · Bansal et al., Does the whole exceed its parts? (CHI 2021)
- NN/g: AI Hallucinations: What Designers Need to Know · Explainable AI in Chat Interfaces
Books, courses and videos to go deeper
Related
Agentic design · Conversational and generative UI · Evals for designers · Trust calibration · Confidence indicators · Explainability · AI error states · Automation bias