02 / Check the result
How do you check an AI answer?
Keep the source beside the answer. Check the claims that matter, look for missing information, and separate what you verified from what still needs checking.
Asking AI “Are you sure?” can start another review. Its reassurance isn’t evidence that the answer is right.
What are you checking for?
Start with the intended use. Before sharing a meeting summary, check the decisions and commitments. Before using research, open the sources and check whether they support the claims.
Write down what would make this particular answer unusable. That gives your review a target.
- Supported: can you find the claim in the original material?
- Complete: did the answer leave out an important condition or disagreement?
- Accurate: do the names, dates, numbers, and units match?
- Clear about gaps: is unknown information marked as unknown?
Can you spot the two mistakes?
This fictional example contains two deliberate mistakes. Read the source before opening the check.
Original meeting note
“Maya will send the revised slides on Thursday. We need to ask the venue about step-free access; no owner was assigned. Leo suggested a rehearsal, but we didn’t agree to one.”
Draft to review
- Maya: send the revised slides on Thursday.
- Maya: ask the venue about step-free access.
- Leo: organise the rehearsal.
Show the check
The venue task has no owner. The draft assigned it to Maya without support. Correct it to “Owner not assigned.”
The rehearsal wasn’t agreed. Move it out of the action list and label it “Suggestion to discuss.”
The first action is supported. Thursday is stated; don’t invent a calendar date unless the meeting context establishes which Thursday.
A useful correction preserves the gaps. It doesn’t replace one guess with another.
How do you check AI research and calculations?
For AI research, open the cited sources and compare each important claim with the relevant passage. Check whether the source’s date and scope fit the claim. A working link alone doesn’t establish support.
For arithmetic, recalculate using the original numbers and units. For a tool or workflow, try a small test with an expected result. If you don’t have the knowledge to judge the answer, get someone who does to review it before relying on it.
Another AI can help find possible errors. Agreement between models still needs evidence behind it.
What should an AI verification record include?
Record the claim, its source or independent test, what the evidence shows, and any correction or uncertainty. That gives another person enough context to follow your review.
Result I plan to use: Claim or action checked: Original source or independent test: What the evidence shows: Correction needed: Still uncertain: My decision: use / revise / ask for helpDownload the checklist
Further reading
NIST’s research examines whether a source backs up an AI-generated claim and whether important details were left out. The checklist here applies that evidence-first approach to everyday work.
NIST: Building Evaluation Probes into Agentic AI ↗Hrishi Bhat builds AI workflows and teaches non-technical professionals how to use them, with results they can check.