Designing for Trust Indicators in AI-Powered Features

On an AI-assisted document review tool we built for a legal-adjacent client, early testers either over-trusted flagged clauses (treating an AI suggestion as fact) or ignored the feature entirely after one wrong flag. Neither extreme was safe. We addressed it by attaching a visible confidence signal to every flag — not a raw percentage, which users didn't know how to interpret, but a three-tier label: 'likely,' 'possible,' 'worth a second look.'

Source attribution turned out to be the single highest-leverage trust feature we've shipped across multiple AI products. Even a small 'based on section 4.2' link next to a generated summary changed how users interacted with the output — they stopped treating it as a black-box answer and started treating it as a starting point they could verify. Usage of the underlying document viewer actually increased after we added citations, which we read as a sign users were engaging more critically, not less.

We also had to design for the AI being wrong, not just for the AI being right. A feedback control — a simple thumbs up/down or 'not quite' button — needs to sit directly next to the output, not buried in a settings menu, because the moment a user notices an error is the only moment they're motivated to report it. On one client project, moving this control from a separate feedback tab to inline with each result increased correction submissions by several times over.

The interface state for 'the model isn't confident here' needed to look meaningfully different from a normal result, not just carry different text. We use a visibly lighter visual treatment — dashed borders, muted color, an explicit icon — for low-confidence AI output, reserving the fully styled, high-contrast treatment for content a human has verified or the model is highly confident about. Making uncertainty visually distinct, rather than just verbally caveated, is what actually changed user behavior.