What would make you trust an AI assistant more: clear sources, better explanations, visible limits,…

Trust should be earned by making the assistant checkable, not by making it sound confident.

My minimum requirements would be:

1. sources shown next to claims, with dates and scope;
2. a clear distinction between retrieved evidence, inference, and generated text;
3. visible limitations and uncertainty;
4. a way to inspect the inputs, tool calls, and output before an external action;
5. human approval for high-impact or irreversible actions;
6. logs and a correction path when it gets something wrong.

NIST’s AI Risk Management Framework lists validity and reliability, safety, security, accountability, transparency, explainability, privacy, and fairness as separate characteristics of trustworthy AI. That is a useful warning against reducing trust to “the model explains its answer.” An explanation can be fluent and still be wrong.

I would also separate trust in a single answer from trust in the system. A cited answer may be checkable, but the system still needs sensible data retention, permission boundaries, monitoring, and a reliable way to stop or undo actions. The most trustworthy default is not maximum autonomy; it is maximum useful visibility at the moment a person must decide whether to accept the result.