Put evidence to work
How to verify AI answers: a practical checklist
Verify the claims that matter to your decision, using evidence you can inspect. Keep a record of what you checked and what you still do not know.
The six-step verification checklist
1. Break the answer into claims
Separate factual assertions from calculations, interpretations, suggestions, and preferences. “This software saves 40% of time and is the best option for us” contains a measurement claim and a recommendation. They need different kinds of support.
2. Decide what needs checking first
Start with claims that would change your choice or be costly to get wrong. For a draft invitation, tone may be the main concern. For a procurement decision, price, compatibility, terms, and measured performance deserve direct checks. A checklist is not a substitute for qualified review when an error could cause serious harm.
3. Find the original source
Open the cited document or locate the original record independently. Prefer the organization’s own documentation for its product, the original study for a research claim, or your own dataset for an internal result. Search snippets and AI summaries can point you toward evidence; they are not replacements for reading it.
4. Check the match
Find the exact passage, table, or calculation supporting the statement. Compare the date, population, version, units, and conditions. A real source can be attached to the wrong claim. A source describing one team does not automatically establish a result for all teams.
5. Recalculate and look for contrary evidence
Check arithmetic with a calculator or spreadsheet using the original inputs. Distinguish percentages from percentage points, and absolute changes from relative changes. Look for relevant limitations or contrary findings rather than collecting only confirming examples.
6. Record a bounded conclusion
Use supported when the inspected evidence supports the exact claim, contradicted when it conflicts with it, and unresolved when the evidence is missing, mixed, outdated, or insufficient. “Supported” describes your current review, not a permanent guarantee.
How to check an AI-generated citation
- Locate the title and authors on the publisher’s site or in a reliable scholarly index.
- Check the year, journal, and identifier against the record.
- Open the source and locate the relevant passage.
- Distinguish the source’s actual finding from the AI’s interpretation.
- If you cannot access enough of the source, record that limit. Do not claim to have verified its contents.
A citation that looks professional may be invented. A citation that exists may still be irrelevant. If a search finds nothing, mark it unresolved while investigating; absence from one search is not conclusive proof that it does not exist.
For a case where the paper exists but the claim goes too far, follow the worked guide to verifying AI-generated citations.
Worked example: “40% faster” becomes a narrower claim
Fictional training example: an AI draft says, “The pilot proves the new workflow improves every team’s productivity by 40% with no quality loss.” The supplied pilot note says: “Ten tasks in one team. Mean completion time: 50 minutes before, 30 after. Quality was not measured.”
| Claim | Check | Result |
|---|---|---|
| Mean task time fell by 40% | (50 − 30) ÷ 50 × 100 = 40% | Supported within the example |
| Every team gains the same benefit | Only one team was observed | Unresolved |
| Quality was measured and unchanged | The note says quality was not measured | Contradicted |
| There was no quality loss | No quality data are available | Unresolved |
Revised conclusion: “In a ten-task pilot with one team, mean completion time fell from 50 to 30 minutes, a 40% reduction. Quality and results in other teams remain untested.”
The example also shows why “not established” is different from “false.” The other teams might benefit, but this note cannot establish that conclusion.
A prompt that makes verification easier
Break your answer into factual claims, assumptions, and recommendations. For each factual claim, identify the original source and relevant passage if available. Distinguish material you actually checked from material you did not check. List missing evidence and avoid inventing citations.
This prompt organizes the work; it does not certify the response. Open the sources yourself. See Prompt Engineering for more help framing the task.
Does asking a second AI verify the first answer?
No. A second response may expose an inconsistency or suggest a source, but two answers can share the same unsupported claim. Use their disagreement or agreement to guide investigation, then inspect the evidence.
Record your checks
The free worksheet stores a claim, a source reference, evidence notes, and your assessment. Its review button highlights incomplete records; it does not browse sources or decide what is true.
Open the verification worksheetFor background, read why AI sounds confident and fluency versus evidence.