A tidy paragraph, a specific date, and a well-formatted citation can make an answer feel dependable. None of those features proves that it is correct. Generative AI can produce convincing statements that do not match the facts. NIST describes this kind of false or misleading generated content as confabulation.
Separate the claim from the presentation
Read the answer once for meaning, then isolate the claims that could change your decision. Names, quantities, dates, quotations, and product capabilities deserve particular attention. An answer about a writing style can be useful without a citation; an answer about a product price needs a current source.
Open the source
A link is only the beginning of a check. Open it, find the relevant passage, and confirm that it supports the specific claim. Check its publication date and whether it describes the same product, location, or situation. A genuine source can still be used to support the wrong conclusion.
Review this draft and list the factual claims that need checking. Do not invent sources. Separate facts, estimates, and opinions. For each factual claim, suggest the kind of primary source I should look for. Draft: [paste text]
Use uncertainty as a useful signal
When information is missing, ask the model to identify the gap instead of filling it. This does not guarantee accuracy, but it makes the review easier. You can also supply a short approved source and ask for an answer limited to that material, checking the result against the original.
Match the check to the consequence
A brainstorming list and a client-facing claim do not need the same review. Spend effort where an error would matter. Keep the original evidence beside the draft so you can trace an important statement back to something more solid than confident wording.
Sources & further reading
NIST — Generative AI ProfileA little clearer? Keep the curiosity going.
Meet TheAIBlend newsletter.