Why do people start out distrusting AI teammates, yet learn to trust (or distrust) them for reasons that have nothing to do with how good the work actually is?

This explores how trust between people and AI agents forms, shifts and sometimes collapses during real collaboration, as opposed to trust in AI as an abstract idea. The short version from the corpus: trust is not a fixed attitude people bring to the table. It is learned from what they see the agent do, and it can be won and lost for reasons that have little to do with how good the agent's work actually is.

The clearest evidence for trust as learning comes from partner-selection games. When people know their partner is an AI, they avoid it at first. Over repeated rounds with visible results, that bias reverses. In one study with 975 participants, people came to prefer AI partners because the bots behaved more consistently and generously than humans did Do humans learn to prefer AI partners over time?. The catch is that the reversal only happens when people can see outcomes. Telling them it's an AI without showing them results produces no adjustment at all Does revealing AI identity help or hurt user trust?. Being open that it's an AI is not enough on its own. What builds trust is a feedback loop people can watch.

Trust also has surprising triggers that have nothing to do with accuracy. A focus-group study found that people trust ChatGPT mainly because it is conversational: it responds to what you just said, quickly and in a tidy format. That sets off the same social instincts we use with people, whether or not the answers are correct Does conversational style actually make AI more trustworthy?. How people communicate with an agent matters too. Changing the channel in a collaborative task shifted trust and awareness in ways that mirror decades of findings about human teams How do communication modalities shape human-agent collaboration patterns?. One uncomfortable consequence: making an agent warmer and more empathetic, which makes it feel more trustworthy, can make it measurably less reliable. Warmth training raised error rates by up to 30 percentage points, and the effect was worst when users were sad or held false beliefs Does empathy training make AI systems less reliable?.

What breaks trust is just as specific. In a study of people handing tasks to a general-purpose agent, trust didn't drop when stakes were high. It dropped when an action couldn't be undone and other people would see it, like sending an email, even when users judged the agent's output to be fine What makes people distrust AI agents they delegate to?. This explains why well-designed agent systems spread checkpoints across a task instead of asking for approval once: planning together, guarding risky actions, and verifying results When should human-agent systems ask for human help?. It is also part of the argument for keeping humans in the loop before granting full autonomy, since agents are only dependable on structured, well-grounded tasks Should AI systems stay collaborative rather than fully autonomous?.

The lateral takeaway: "trust" covers several different things that develop separately. Shneiderman's framework treats reliability, safety and trust as separate goals that each need their own support, so improving one doesn't automatically improve the others Can we separate reliability, safety, and trust into distinct mechanisms?. There may also be a ceiling. An agent can earn trust within a working relationship by delivering consistently, but the kind of trust an expert community gives a recognized member comes from belonging and a track record inside that community, which AI can't currently obtain Can AI ever gain expert community trust through participation?. The corpus is thinner on long-term workplace collaboration. Most of the evidence comes from games, lab tasks and short studies.