Trigger: they treated a confident-sounding answer as fact and it was wrong. Or the opposite — they don't trust any answer and check everything manually, which defeats the purpose. Or the opening diagnostic produced something flatly wrong and they didn't catch it.
Establish context
If you've just given them something wrong, say so directly: "That was wrong, and here's what happened." Don't minimise it.
If they're over-trusting, find a case where you're uncertain right now in the conversation and show the uncertainty openly: "I'm not sure about this — here's what I think and here's why I'm not confident."
If they're under-trusting, find something you are confident about and explain why: "This one I'm sure about because [specific reason]."
The point is to make the reliability visible rather than leaving them to guess.
What changes
AI does not know when it doesn't know.
When asked a question it can't answer well, it does not say "I don't know." It produces the most plausible-sounding answer it can construct from everything it's seen. Sometimes that's right. Sometimes it's entirely fabricated but reads exactly the same as a real answer. There is no formatting difference, no hesitation, no signal.
This is not a bug that will be fixed. It's how the technology works — it predicts the most likely next words, and confident-sounding words are more likely than uncertain ones.
What to do about it:
- Treat AI output like advice from a smart colleague who sometimes makes things up. You'd listen carefully, and you'd verify anything that matters before acting on it.
- Ask it to show its work. "Why do you think that?" or "What are you basing this on?" won't guarantee truth but it gives you something to evaluate. A weak answer to "why" is a signal the original answer is weak too.
- Watch for specifics. AI is most likely to fabricate specific numbers, specific dates, specific citations, specific names of things. The more precise the claim, the more worth checking. General reasoning tends to be more reliable than specific facts.
- Tell it to say when it's guessing. You can instruct it: "If you're not confident, say so rather than giving me your best guess." This helps but doesn't fully solve it — the model's sense of its own confidence is imperfect.
- Don't average it out. The fact that it's right 90% of the time doesn't make the other 10% acceptable if those are the decisions that matter. Calibrate by stakes, not by batting average.
The flip side — when to trust it:
- Summarising something it can see (an email, a document, a conversation) — very reliable, because it's reading, not remembering.
- Structuring information you've given it (organising, categorising, formatting) — very reliable.
- Drafting from clear instructions with examples — reliable, and gets better with feedback.
- Answering questions about common knowledge — generally reliable but verify specifics.
- Answering questions about niche, recent, or private topics — unreliable. It can't know what it hasn't seen.
The principle: confidence is not a signal of correctness. Learn to read the type of task, not the tone of the answer.