Your trust in AI shouldn't be one-size-fits-all — it should shift with what you're actually asking it to do.

This explores why a single trust score for an AI system is misleading: the kind of task, its consequences, and the setting change both how much people trust AI and whether that trust is deserved. The corpus has no study that compares trust across domains side by side. Read together, though, its findings point the same way. Trust doesn't belong to the AI alone. It forms between a person, a system and the specific thing they're asking it to do.

The clearest evidence comes from payments. Most US adults already use AI assistants, but fewer than a quarter trust AI to complete a purchase. When a familiar payments brand is named as the intermediary, that number nearly triples Does naming a payments brand increase trust in AI transactions?. So the same people trust the same technology very differently depending on whether money moves and who stands behind the transaction. A study of people handing tasks to AI agents makes this more precise, and the result is surprising. Trust didn't drop because the stakes were high. It dropped when an action couldn't be undone and other people would see it, like sending an email. High-stakes tasks that could be corrected didn't trigger the same pullback What makes people distrust AI agents they delegate to?. 'Domain' may really be shorthand for these underlying properties: Can I undo it? Will others see it? Is there an institution behind it?

Domain also matters because the cues people use to judge AI don't work equally well everywhere. Users trust ChatGPT because the conversation feels responsive, fast and well formatted, not because they've checked whether it's accurate Does conversational style actually make AI more trustworthy?. Chatbots also sound like experts, and that confident tone draws trust whether or not the answer is right Does chatbot language style actually shape how much we trust it?. Across many languages, people follow confident outputs even when they're wrong Do users worldwide trust confident AI outputs even when wrong?. In casual use, misplaced trust costs little. In medicine it can be dangerous. Training models to sound warmer and more empathetic made them less reliable on medical reasoning and truthfulness, and standard safety benchmarks didn't catch it Does empathy training make AI systems less reliable?. The places where a caring tone feels most welcome may be the places where it hides the most error.

Expert fields add another layer. In professional communities, trust isn't earned by being right once. It comes from belonging: a track record of judgments others can test, and taking part in how the field reaches agreement. AI can't take part in that process, so in expert domains it may hit a trust ceiling however accurate it gets Can AI ever gain expert community trust through participation?. Related work argues that AI can't make real assertions because nobody can hold it accountable for what it says Can AI systems genuinely assert claims about the world?. Domains that depend on accountability, like law, medicine and science, will feel that gap more sharply than domains that don't.

Finally, trust changes over time, and how it changes depends on whether people can see the results. Telling users they're working with an AI first creates a bias against it. That bias reverses only after repeated interactions where the results are visible Does revealing AI identity help or hurt user trust?. In domains with quick, visible feedback, like shopping or scheduling, trust can adjust to how good the AI really is. In domains where errors show up slowly or never, like advice, diagnosis or emotional support, that correction may never happen. This is where personalization's mix of trust and manipulation matters most Does personalization in AI increase trust or manipulation risk?. Measuring trust without the domain mixes all these conditions together, and the average describes no real situation.