Sounding sure is not evidence
Why confident errors happen
A model produces the next piece of text that fits, not one it has checked against the world. So a smooth, confident, wrong answer is an expected outcome of the design. People call this hallucination. It is most likely with specific details: dates, figures, quotes, names, references and links. A made-up book title or court case can look exactly like a real one, which is why checking matters most for the details you plan to use.
Pull out the exact claim
Answers mix three kinds of statement: claims (something is so: a date, a number, an event), explanations (why it happened) and opinions (what someone feels). You fact-check the claims. Rewrite each one as a single testable line: "The Eiffel Tower was completed in 1889." Vague claims cannot be checked; sharpening them often reveals that the answer said less than it seemed to.
Match the source to the claim
A source is not reliable in general; it is suited, or not, to a particular claim. For an opening date, the monument's own history fits well. For a medicine dose, the official leaflet or a pharmacist. For a law, the government's legislation site. A random blog or social post may mention the fact but is a weak fit. Check the date too: a 2008 visitor figure says little about visitor numbers today.
Record what the evidence says
After checking, give each claim one of three verdicts: confirmed, contradicted, or unresolved. Two independent good sources agreeing gives real confidence. "Unresolved" is a respectable answer, far better than guessing. You can also ask the AI itself for sources, then open them: invented links and references are common, and a link that does not say what the AI claimed is a contradiction, not a confirmation.
Think of a claim an AI gave you recently. Which part of it would you actually need to check before relying on it?