Confidence is a writing style, not evidence
People naturally treat a clear, detailed answer as a sign that the speaker knows the subject. AI makes that shortcut dangerous. The same system that produces an excellent explanation can also generate a convincing detail that never existed.
NIST uses the term confabulation for confidently presented false or erroneous generated content. It is not a rare formatting bug. It follows from the way generative systems create responses.
Use the original source when the answer matters
Search, uploaded documents, and code execution can improve reliability when used appropriately. They do not remove the need to check. A citation can point to a real page that does not support the sentence beside it.
For money, health, law, safety, customer commitments, or anything difficult to reverse, open the original source. Check the number, date, definition, and surrounding context yourself or with a qualified professional.
Match the review to the risk
Not every output needs the same process. Use a simple risk ladder.
- Low consequence: brainstorms and personal outlines can be reviewed lightly.
- Moderate consequence: public writing, analysis, and work products need source and logic checks.
- High consequence: regulated advice, payments, contracts, and customer commitments require expert or accountable human review.
Ask the system to reveal uncertainty
Useful review prompts include: “Which claims depend on a source?” “What did you infer?” “What information is missing?” and “What would make this answer wrong?” These questions do not guarantee honesty, but they make the review more deliberate.
Take one AI answer you planned to use. Mark every statement as supplied fact, sourced fact, calculation, or inference. Verify the categories with the highest consequence.
Sources and further reading
Product capabilities and policies change. These are the primary sources used for this guide.