Designing for calibrated trust
The goal of an AI interface is not more trust. It is the right amount. Two decades of human-AI research, distilled into the design moves we reach for first.
In 2019 a team at Microsoft Research presented eighteen guidelines for human-AI interaction at CHI, the main conference on human-computer interaction. They distilled more than twenty years of research and practice, then tested the guidelines with 49 design practitioners against 20 AI products people use every day. Google’s People + AI Guidebook reaches the same conclusion from a different direction: people should “know when to trust the system’s predictions and when to apply their own judgement.”
Both bodies of work point at the same idea. Trust in an AI product is a dial, not a goal. Too little, and people ignore a feature that could help them. Too much, and they accept answers they should have questioned. Good design aims for calibration: trust that matches how reliable the system actually is, in this situation, for this task.
Set expectations before the first answer
The first two Microsoft guidelines are the ones most often skipped: make clear what the system can do, and make clear how well it can do it. A blank prompt box does neither. It asks people to guess the system’s abilities, and most guess conservatively, try one thing and leave.
- Offer concrete starting points that show the range of what the feature handles well.
- Say plainly where it is weak: “Drafts are a good start. Check names and figures.”
- Show the shape of a good result before anyone asks, so the first output is not a surprise.
Explain partially, and at the right moment
Explanation is where AI interfaces most often overreach. The PAIR Guidebook recommends partial explanations: show what matters for the decision at hand, not everything the model did. Pointing to the sources a summary drew on, or the two features that drove a recommendation, does more for trust than a paragraph of reasoning nobody reads.
Timing matters as much as content. The Guidebook notes that “the perfect time to show explanations is in response to a user’s action.” An explanation attached to the result someone is about to act on is read. The same explanation in an onboarding tour is forgotten.
Show confidence only when it changes a decision
A percentage next to every answer feels rigorous and often misleads. PAIR suggests showing confidence only where it actually improves decisions, and choosing the form carefully: simple categories like high and low, several alternatives instead of one confident answer, or a visual range. Raw numbers are the riskiest option, because people read 72% very differently from how the model meant it.
Our default is to offer alternatives when the system is unsure. Three candidate answers communicate uncertainty more honestly than one answer with a number attached, and they give people something to do with it.
Design the wrong answer first
A model will be wrong some of the time, so the Microsoft guidelines spend a whole phase on that moment. Support efficient correction: make it easy to edit, refine or recover. Scope services when in doubt: ask a clarifying question, or do less, rather than guessing at a goal. Support efficient dismissal: let people ignore a suggestion without a fight.
In practice this means editing results in place instead of only regenerating them, keeping undo one step away even for actions the AI took, and treating “I am not sure, did you mean…” as a feature rather than a failure.
A trustworthy AI product is not one that is always right. It is one that makes it easy to tell when it is not.
Trust is earned over time, and can be spent in a release
The last group of guidelines is about change: learn from behaviour, update and adapt cautiously, and notify people about changes. AI products change more often than most software. A new model can shift tone, accuracy or speed overnight. That is exactly the kind of change people should hear about from the product, not discover in a worse answer on a Tuesday.
The questions we ask in a design review
- Could a new user tell what this feature is good at within ten seconds?
- Does the interface say how reliable it is, in words, where it matters?
- Can someone see why a result came out the way it did, at the moment they act on it?
- When the system is unsure, does it ask, narrow down or offer alternatives?
- Can every AI action be corrected or undone in one step?
- Will people know when the model behind the feature changes?
