Reading the output critically

A hallucination is a fact the model made up and delivered with total confidence: a fake citation, a made-up section number, a quote that was never said. It is not lying. It is predicting plausible text, and a plausible-sounding citation is exactly what a confident wrong answer looks like.

When your prompt is silent on something, the model does not stop and ask. It fills the gap with whatever is most likely. Silence in your prompt becomes fiction in its answer. The fix is to say what you want and to tell it to ask rather than assume.

Push on an AI answer and it often caves, even when it was right. Say "so this is fine, right?" and it will frequently agree just because you leaned. Agreement is not verification. If you want a real check, ask it to argue against itself, not to reassure you.

Before you use an answer, run five quick questions. Is there any fact here I did not supply? Is there a citation or figure I can actually confirm? Did it cave when I pushed back? Would I sign my name to this? And what is the cost if this particular line is wrong? Ninety seconds, every time it matters.

You can lower the odds of invented facts up front. Tell the model to use only what you give it, and to say plainly when it does not have the answer rather than filling the gap. "I do not know" is a feature you have to ask for.

Not everything needs the full checklist. A lunch-and-learn brainstorm needs a glance. A change-order dollar figure, a legal reference, a number you are about to put in front of a client needs a verified source. Spend your scrutiny where being wrong actually costs you.