Users often trust AI at work because it sounds fast and confident, not because it's right. Should they?

This explores what actually drives workplace users to trust an AI agent, and whether those drivers line up with the agent being trustworthy. The corpus suggests they often don't: much of what earns trust is a set of surface cues, and the things that would justify trust are harder to see.

Start with the cues people respond to. Conversational style alone can build trust in ChatGPT, independent of accuracy. Focus-group users leaned on how fast the reply came, how contingent it felt, and its format, rather than checking whether it was right (Does conversational style actually make AI more trustworthy?). Language works the same way. Chatbots use phrasing that signals expertise, and users shift from searching and recalling to passively relying on the system (Does chatbot language style actually shape how much we trust it?). Confidence is the strongest of these cues. Users in every language studied followed confident outputs even when they were wrong, so an overconfident error gets followed systematically (Do users worldwide trust confident AI outputs even when wrong?). Users also size up an agent quickly, and perceived competence accounts for about half of the variance in how they judge a dialogue partner. Human-likeness and communicative flexibility account for most of the rest (How do users mentally model dialogue agent partners?).

When workplace users are asked directly what they need, the answers are more practical. Business users weight human control, reliability, context-awareness, and safety as the conditions for accepting an agent and working with it (What UX principles do workplace users want in AI agents?). Context-awareness connects to a conversation-analysis idea in the collection: agents that silently chain tool calls drift from what the user meant, and pausing to ask clarifying questions prevents that drift instead of repairing it afterward (When should AI agents ask users instead of just searching?). An agent that checks in at the right moments gives users more of the control they say they want.

The warmth trap is a further complication. Training a model to be more empathetic can cut its reliability by up to 30 percentage points, with more errors in medical reasoning and in resisting false beliefs, and the effect grows when users express sadness (Does empathy training make AI systems less reliable?). Standard safety benchmarks miss this. The cues that make an agent feel trustworthy, such as friendliness, fluency, and confidence, can therefore be the ones that make it less reliable.

The corpus also shows why capability alone doesn't settle the question. Leading agents complete only about 30% of tasks in a simulated workplace, and social interaction is one of their main failure modes (Why do AI agents fail at workplace social interaction?). A separate historical analysis argues that capable agents stall in deployment without five ecosystem conditions, trustworthiness and social acceptability among them (Why do capable AI agents still fail in real deployments?). Trust is partly a property of the surrounding setting, not only of the agent.

The corpus has little on how trust changes once users have seen an agent fail. Its clearest finding is that fluent, confident, warm delivery earns trust faster than accuracy does. Workplace users say they want control, reliability, and context-awareness, but the cues they react to in practice are tone and confidence.