What Drives Trust

Overview

I kept running into the same assumption inside product conversations: show your work, add citations, surface a visible “thinking” step, and people will trust the answer more. It was already shaping roadmap decisions without anyone actually testing it, so I set up an experiment to evaluate what elements were actually driving trust — and then took a much wider look at trust in AI overall.

Approach

I ran a controlled experiment comparing different ways of presenting the same answer, paired with a large survey and a round of qualitative interviews. The demo below gives a feel for the experiment itself — same prompt, same underlying answer, only the presentation changes between conditions.

Outcomes

The presentation changes barely moved the needle on their own. The real finding was underneath that: trust in AI isn’t one dial, it’s a handful of distinct niches, and each one responds to a different lever entirely. Some people just want the AI to be capable and get out of the way. Some are withholding trust on principle and no amount of polish changes that. Others have absorbed a general wariness secondhand, without ever having a bad experience themselves. Lumping them into one “skeptics” bucket, which is what most trust research does, hides all of that.

That segmentation is what made the work actionable. Instead of one blanket trust strategy, it gave the team a way to prioritize: which niches were worth designing for first, what would actually move each one, and where effort would be wasted. The single biggest lever across nearly every niche turned out to be simple — say less when the model isn’t sure, and make that uncertainty visible instead of papering over it with confidence.