Spotting unreliable answers

Watch for claims that need checking #

Treat polished wording as presentation, rather than evidence. Give extra attention to precise figures, quotations, references and confident recommendations that could affect a decision.

• Open references and check author, title, date and the supporting passage.
• Distinguish a real source from a source that actually supports the claim.
• Look for missing conditions, exceptions and conflicting evidence.
• Check whether the answer has assumed facts you did not provide.

A worked example: an unsupported recommendation #

Suppose an answer says a workshop format will improve performance by 40%, but supplies no supporting study. Do not retain the number because it sounds persuasive. Locate reliable evidence or remove the claim and label the expected benefit as something to test.

List the claims in this draft that need verification.
For each, identify the supplied supporting passage, or mark it unsupported.
Separate facts, assumptions and recommendations.
Do not create references to fill gaps.

Verify outside the generated answer #

Asking the same AI whether it is correct is not independent verification. Consult original documents, reproduce calculations and use an appropriate subject expert where the decision requires one. Record what remains unresolved.