A team ships a useful AI feature. Early users love it. Then the model gets something confidently wrong, and the user's relationship with that feature changes -- sometimes permanently. Trust broken by an AI system is harder to rebuild than trust broken by a human, because users have no mental model for recalibration. It seemed certain. It was wrong. That combination is uniquely corrosive.
Four principles that hold
Show the reasoning, not just the result. 'Suggested because you responded quickly to emails with attachments' is more trustworthy than 'Suggested for you' -- even if the underlying logic is identical. The reasoning gives users something to evaluate.
Make correction easy and visible. A system that makes correction easy is implicitly saying: we know we might be wrong, and your judgment matters. That posture builds more trust than any accuracy claim.
Surface uncertainty in the interface, not just in the model. If a model's internal confidence score is not reflected in the UI, users receive every output as equally certain. Softer language, hedging labels, probability cues -- these belong in the design system, not as an afterthought.
Distinguish assistance from authority. The clearest trust failures come from interfaces where the AI's role was ambiguous. Was it making a decision or a suggestion? These are design questions. The model cannot answer them.
On full automation
Fully automated AI actions require graduated trust: the system starts by suggesting, then drafting, then acting with review, then acting with notification. The user controls which level they are at. Trust extended at the user's own pace holds far better than trust asked for upfront.
The products that build lasting relationships around AI are not the ones with the most capable models. They are the ones where users feel in control of a capable tool. That feeling is entirely a design problem.