Practical guides
Verify an AI answer
Treat an AI answer as a collection of claims to examine. Confidence in the wording is not evidence. The amount of checking should depend on the consequences of an error, not on how polished the answer looks.
Method
1. Separate claims
Mark facts, estimates, opinions and recommendations. Prioritise names, dates, figures and statements that could change a decision. Write down the exact claim before searching.
2. Open original evidence
Follow the source and read the supporting passage. Check author, date, country, units and exceptions. A real document can still fail to support the claim attributed to it.
3. Record the decision
Use a small table: claim, source, supporting passage and status. Mark unconfirmed statements as unknown. Correct contradictions before sharing the result; ask a qualified person when the stakes require it.
Example to adapt
Review this explanation of a historical event. List three checkable claims, identify the type of original evidence needed and leave the status unknown until each source has been read.
Checklist
Checks stay on this page; ticking a box does not automatically validate the result.
Deliverable
A claim register with confirmed, contradicted and unresolved statements.
Limitations
Asking a second model is another generated opinion, not independent evidence. Do not turn an unresolved point into a definitive conclusion.
Sources for further reading and checking
These independent resources help you find a reference. Check the language and scope: a French resource may not describe rules in another country. These are not sources automatically consulted for generated answers.
- NIST — Generative AI Profile (AI 600-1)
A generative AI risk profile for identifying and managing risks.
Language: en · Scope: US
Link check: 2026-10-06 · These links are reading resources. The answer service does not consult them automatically. A checked link does not guarantee every document is current.