Does trust in an AI advisor track its actual accuracy, or do other cues matter just as much?

This explores whether trust in an AI advisor follows its track record, or whether something else is doing most of the work. The collection's short answer is that past performance matters, but only when people can actually see the results, and even then it competes with signals that have nothing to do with accuracy. None of these studies measures the size of the effect directly. Together, though, they show when a track record counts and when it gets ignored.

The clearest evidence that track record matters comes from studies where people are told they're working with an AI. Does revealing AI identity help or hurt user trust? found that people first avoid an AI partner once they know it's an AI, but that this bias reverses after repeated rounds where they can see the outcomes. The surprising part is that being told it's an AI, without any feedback on results, didn't help people calibrate at all. Performance only builds trust when it's visible and repeated. Expertise works the same way in human communities: Can AI ever gain expert community trust through participation? argues that experts earn authority through a judgment history that others can test over time. That note argues AI can't take part in that process, so being accurate in a single case doesn't turn into standing.

In practice, trust often comes loose from accuracy. Does conversational style actually make AI more trustworthy? found that ChatGPT users based their trust on how the exchange felt (responsive, fast, well-formatted) more than on whether the answers were reliable. How much should we trust AI-generated data in inference? describes a similar pattern in research workflows: people treat AI output as fully trustworthy by default because it sounds confident, not because it has earned that trust. Does personalization in AI increase trust or manipulation risk? adds that memory and personas can build trust through rapport, and the same mechanisms can be used to manipulate. So a smooth advisor can collect trust that its track record doesn't justify.

Trust can also drop even when performance is fine. In What makes people distrust AI agents they delegate to?, people pulled back from an AI agent on tasks that couldn't be undone and that others would see, such as sending an email, even when they rated the output as adequate. Tasks with high stakes that could still be corrected didn't trigger the same drop. In these moments, what decides trust is the cost of being wrong, not how often the AI was right before.

The idea you may not have expected to want: the setup of the help can matter as much as the track record. Can AI guidance reduce anchoring bias better than AI decisions? suggests that when an AI points out what to look at instead of handing over a verdict, people stop anchoring on its answer, so a good or bad record matters less. The open question this collection can't yet answer is a number: how many visible successes does it take to offset one visible, irreversible failure? If you find research on that, it would fill a real gap here.