An AI interface should help users judge when its output deserves reliance. Fly.io's design essay applies that idea to the boundary between a suggestion the user can reject and an action the system performs. Showing proposed changes and explaining operating limits are ways to make that boundary visible.
The original research survey explains why increased trust alone is an inadequate success measure. A test must connect users' reliance with the system's actual capabilities to identify over-trust or under-trust. Feedback cues need testing too: the survey discusses cases where added explanations or uncertainty indicators themselves encouraged misplaced reliance.
Terms in this piece · Glossary
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
