Verify an AI-generated answer
Break an AI response into claims, open its cited sources and document what remains uncertain.
Follow the method ↓
The answer in 30 seconds
A fluent answer can combine sound advice with invented references, outdated instructions or unsupported details. The level of checking should match the consequences of using it.
Extract the specific claims you might use, inspect each cited source and its date, verify numbers and quotations in the original, and leave unsupported claims out of your final decision.
Examples to adapt
Find the original table and compare population, year and unit.
Check current official instructions for your device and version.
Use a qualified human or competent authority when an error could affect rights, health or money.
Follow the method
- 1Separate the claims
Mark dates, quantities, quotations and instructions that matter.
- 2Open the evidence
Visit the actual cited page; a plausible URL or source title is insufficient.
- 3Check fit
Compare the source's scope, date and wording with the answer.
- 4Record limits
Keep confirmed parts, label uncertainty and seek expert help for high-stakes use.
Workshop: audit three claims in an answer
Highlight one number, one attribution and one recommendation. For each, find an independent source, open it and locate the precise supporting passage. Record the source date and any conditions limiting the claim.
A fluent answer can combine correct facts with links that prove nothing. If one claim fails, rewrite the conclusion and identify what remains unverified.
A closer look
Check each claim at the right level
Fictional example: an answer states “120 facilities open in 2026, up 20%, thanks to a new policy”. This sentence contains three different types of claim.
| Claim | Evidence to seek | Limit to retain |
|---|---|---|
| 120 facilities open in 2026 | Dated table, definition of “open” and geographic scope | An address list alone does not establish public access |
| 20% increase | Comparable baseline of 100: (120 − 100) ÷ 100 | Scope and definition must remain consistent |
| Thanks to a policy | Evaluation capable of examining causation | Sequence in time alone does not demonstrate cause |
You can confirm a count, correct a rate and leave the cause unresolved. Checking does not require accepting or rejecting the entire sentence together.
Does a link supplied by AI lead directly to evidence?
Open the source and locate the exact passage. Check that it exists, matches the scope and supports the claim; a similar title or homepage is insufficient.
A checklist to keep
Use these checks to record what you found. The grid supports a decision; it does not make one for you.
| Check | What to examine | Action |
|---|---|---|
| Claim | What exactly would you repeat? | Extract |
| Citation | Does the linked page exist? | Open |
| Support | Does it prove this wording? | Compare |
| Date | Is it applicable now? | Check |
| Consequence | How harmful would error be? | Escalate |
Read the checklist directly in the table.
What to check
A citation can exist without supporting the adjacent claim.
Repeated AI answers are not independent confirmation.
Do not paste confidential information into a consumer tool without a clear basis.
Common questions
Does a source list make an answer trustworthy?
No. Each source must be opened and matched to the particular claim.
Can AI still help research?
Yes. Treat it as a way to formulate questions and find leads, then verify the result independently.
Sources and documentation
Documentation consulted on . Examples are illustrative; interfaces and results may change.
Resolve the next uncertainty
Two articles can rely on the same evidence. Group repeated reports by original document before presenting agreement as corroboration.