When people distrust AI as soon as they learn it's AI, can seeing its actual results win that trust back?

This explores whether the trust penalty AI pays when it identifies itself can be worked off over time, simply by people seeing what the AI actually delivers. The corpus says yes, with one important condition. In studies of people choosing partners, participants avoided partners labeled as AI at first, but that preference reversed after repeated rounds where they could see the results Does revealing AI identity help or hurt user trust?. The condition is in how the reversal happens. Repetition alone doesn't repair trust. Visible outcomes do. Disclosure without feedback leaves the bias where it started. So the penalty isn't a fixed verdict about AI. It's a guess people make before they have evidence, and evidence replaces it.

That changes the case for disclosing AI use at all. If the penalty fades once people see results, then hiding AI use to avoid it looks like a bad trade. Schilke and Reimann found that AI use kept quiet and discovered later produces a steeper drop in trust than being upfront Does hidden AI use cost more trust when exposed?. Put the two findings together and disclosure costs you early, while concealment risks a larger loss later that you may not recover from. Being open about AI use looks like the better long-term strategy.

The corpus also raises something the question doesn't ask: the trust penalty may be doing useful work. People told that an AI wrote something become more critical and look harder at it, though a large share (34–62%) are still persuaded Does telling people an AI wrote something actually stop them from believing it?. If repeated good outcomes wear down the penalty, they probably wear down that extra scrutiny too. That's where the risks of trusting AI outputs you shouldn't come in: confusing a fluent answer with the facts, mixing up gut feel with reasoning, and having your existing beliefs confirmed back to you Why do people trust AI outputs they shouldn't?. A run of good results is exactly when those habits take hold. So repairing trust and calibrating trust aren't the same thing. Feedback can overshoot.

One alternative in the corpus avoids that overshoot: change what the AI offers instead of asking people to trust its answers. In the Learning to Guide approach, the AI points out the parts of a case worth paying attention to rather than handing over a decision, which reduces people's tendency to anchor on the AI's answer Can AI guidance reduce anchoring bias better than AI decisions?. There's also a ceiling on what any amount of good results can earn. Expert trust is granted through belonging to a community and a track record judged by peers, not through accuracy alone, and AI can't take part in that process Can AI ever gain expert community trust through participation?. Seeing outcomes can turn an individual's suspicion into reliance. It doesn't seem able to give AI the kind of trust that comes from being accepted as an expert.

The practical upshot: the disclosure penalty can be repaired, but only when people can actually see the results. And the repair works best when it builds calibrated trust rather than simply removing the doubt. The corpus is strong on the reversal and on the cost of concealment. It's thin on how long the repaired trust lasts and on what happens after a visible failure. A single bad outcome might bring the penalty right back, and the collection doesn't yet show whether it does.