The trust gap: why users abandon accurate AI
A pattern we keep seeing: AI features get abandoned not because they're wrong, but because users can't tell when they're right. Notes on the trust gap and how design closes it.
We keep running into the same pattern across products: an AI feature is accurate enough to be useful, and users abandon it anyway. When we dig in, the reason is rarely the model. It's that the product never gave people a way to know when to trust the output — so they defaulted to not trusting any of it.
Accuracy is invisible; doubt is loud
Users don't experience your model's accuracy as a number. They experience individual outputs, one at a time, with no error bars. A tool that's right 92% of the time feels untrustworthy if the 8% arrives with the same confident tone and no way to catch it. One bad output early, with no way to see it coming, and the user quietly downgrades the whole feature to "nice but I double-check everything" — which is another way of saying they've stopped using it.
The gap is between "correct" and "verifiable"
There are two different properties a piece of AI output can have: being correct, and being checkable. Teams obsess over the first and ignore the second. But from the user's seat, uncheckable-and-correct and uncheckable-and-wrong feel identical in the moment — both are a leap of faith. Closing the trust gap means making outputs verifiable: showing the source, the reasoning, the data it drew on, or simply flagging when confidence is low.
What we've seen work
The interventions that move trust are unglamorous and consistent. Surfacing a source next to a claim. Letting the user edit rather than accept-or-reject. Making low confidence visible instead of hiding it. Keeping an undo within reach. None of these make the model better; they make its reliability legible, which is what users actually respond to.
Why this compounds
Trust isn't a one-time gate — it's the input to a loop. A feature people trust gets used more, which generates more signal, which makes it better, which earns more trust. A feature people don't trust gets abandoned after one bad experience, and no model upgrade wins those users back, because they've stopped looking. Which is why we treat the trust gap as a first-order product problem, not a polish item.
This is an ongoing thread for us — closely tied to our work on designing AI you can trust and on putting AI inline rather than in a tab.
If you want a concrete read on where your AI feature builds or breaks trust, the free AI Experience & Trust Audit scores it and hands you specific fixes.
More research
Product judgment is the new moat
As models commoditize, the defensible edge moves from the technology to the decisions around it. Notes on why product judgment — not the model — is becoming the moat.
Read →Research note · 5 minAI-as-a-tab is a dead end
The bolt-on chatbot ships fast and demos well, then goes unused. A note on why AI belongs in the flow of the work — and what changes when you move it there.
Read →