This is the honest problem at the centre of using AI at work, and most courses walk straight past it. You asked the question because you did not know the answer. The reply is fluent, structured and entirely plausible. Now what?
"Just verify it" is not advice. If you could verify it quickly you would not have asked. What you need is a small set of moves that raise your confidence without costing more time than the task saved.
Sort the answer before you check it
Most replies are a mix of two things: work done on material you supplied, and claims that came from the model. The first is usually reliable and cheap to spot-check. The second is where fabrication lives. Read it once and mark which is which — the checking work collapses onto the second pile.
Four moves that actually help
- Ask for the source, then go and look. Not "is this true?" — asking the same model to grade itself mostly gets you a confident yes. Ask where the claim comes from, then open that page yourself. A source that does not exist, or that exists and does not say it, is the most common tell there is.
- Ask the same question again in a fresh conversation. Genuine knowledge repeats. Fabrications tend to wobble: the figure moves, the name shifts, the citation changes. Instability is a warning even when you cannot check the underlying fact.
- Recalculate every number by hand. Every one that leaves your desk, not a sample. Models present arithmetic with total composure and get it wrong often enough that spot-checking is not protection.
- Ask what would make the answer wrong. "What assumptions is this resting on, and which of them would change the conclusion?" surfaces the load-bearing assumption you did not think to question — and that question the model is good at.
One check does not work at all: asking the model whether it is sure. It will apologise convincingly when you push back on a correct answer, and defend an incorrect one just as convincingly. Confidence is not a signal in either direction.