Trust is designed, not marketed
Teams often try to earn trust with copy: "powerful", "smart", "accurate". Users ignore it. They form trust from what they see the product do, over and over, in small moments.
Google's People + AI Guidebook frames this well. It organizes its design patterns around the questions a team has to answer, including whether AI adds value at all and how to set the right expectations. Trust starts with those two decisions, before any interface is drawn.
Design principle“The goal is not maximum trust. It is calibrated trust: users who rely on the AI when it is right and double-check it when it is not.”
Set expectations before the first prompt
People arrive with a mental model borrowed from search, chatbots, or science fiction. If your feature does not match it, they will be disappointed on the first try and rarely come back.
- Say what the feature is for in one plain sentence, next to the input.
- Show two or three example inputs that work well, and one that does not.
- Name the limits up front: what data it can see, what it cannot do, and how fresh its knowledge is.
These three lines cost almost nothing to build. They prevent most of the early frustration that teams blame on model quality.
Show the reasoning at the moment it matters
Transparency does not mean dumping model internals on the screen. It means giving users the one fact they need to judge this particular output: where it came from, how sure the system is, or what it assumed.
| Moment | What to show | Example |
|---|---|---|
| Answer from your data | The source it used | A link to the record or document |
| Low confidence | A plain flag, not a percentage | "I am not sure about this date. Please check." |
| An assumption was made | The assumption, editable | "Assuming the last 30 days. Change range" |
| Action about to run | A preview of the effect | "This will email 42 customers" |
Make correction cheap
NN/g observed that people using generative AI rarely accept the first result. They iterate across several steps, because the system seldom delivers exactly what they meant. Design for that loop.
Every output should have a fast way to edit it, regenerate it, undo it, or throw it away. Keep the user's original input visible, so they can adjust it instead of starting over. A feature that is easy to correct feels safe. A feature that is hard to correct feels like a risk, however accurate it is.
A short trust checklist
- Does the user know what this feature can and cannot do?
- Can they see where an answer came from?
- Is uncertainty shown in words a person would use?
- Can they undo any action the AI takes?
- Do you measure when users override the AI, and learn from it?
If you can answer yes to all five, you are ahead of most AI products on the market.



