Verification checklist
Check AI answers before they become decisions.
AI can produce fluent answers quickly, but fluent is not the same as correct. This checklist gives readers a repeatable review process for low-risk daily tasks and a warning system for higher-risk situations.
The seven-step AI output check
- Restate the task. Write down what you actually asked the AI to do. A vague task usually creates a vague answer.
- Separate facts from suggestions. Facts need sources. Suggestions need judgment. Do not review both in the same way.
- Look for missing context. Ask what the answer assumes about time, location, audience, rules, budget, or risk.
- Check names, dates, numbers, and links. These are common failure points because AI may sound confident even when details are wrong.
- Compare with at least one trusted source. Use official documentation, original reports, or reputable references when accuracy matters.
- Review privacy and permission. Make sure the output does not reveal private details or reuse content you should not publish.
- Decide who is responsible. If the answer affects money, health, legal issues, safety, school work, or public claims, a person must own the final decision.
Quick scorecard
| Question | Good sign | Warning sign |
|---|---|---|
| Is the answer specific? | It explains limits and gives context. | It gives broad claims with no details. |
| Are sources needed? | Important claims point to verifiable sources. | It invents citations or avoids evidence. |
| Can you review it? | You understand enough to judge the result. | The topic is outside your knowledge and high risk. |
| Is private data involved? | Names and sensitive details were removed. | The prompt included personal or confidential records. |
Example: checking a summary
If AI summarizes a long article, do not only ask whether the summary sounds good. Check whether it includes the main argument, misses important limitations, changes the author's meaning, or adds claims that were not in the original. A useful follow-up prompt is: "List three claims in your summary that I should verify against the original text."
When not to use AI output as-is
Do not publish or act on AI output without review when the answer affects medical care, legal obligations, financial decisions, hiring, grading, safety, or someone's reputation. In those cases, AI can help organize questions, but it should not be treated as the final authority.
Three common review examples
If AI drafts an email to a customer, check the promise, deadline, tone, and facts before sending. A polite draft can still make a commitment your business cannot keep.
If AI explains a technical topic, ask it to separate core concepts from examples. Then verify the core concept against documentation or a trusted learning source. Examples can be simplified, but the concept should not be distorted.
If AI analyzes a spreadsheet, ask which columns it used, which rows may be missing, and what assumptions it made. A chart or summary can look precise while still being based on incomplete data.
Red flags that need human review
- The answer gives exact numbers without saying where they came from.
- The answer cites a source that cannot be opened or does not support the claim.
- The output changes a policy, contract, assignment, or medical meaning.
- The answer is very confident but ignores important limits or exceptions.
- The prompt included private information that should not be reused or published.
Copy this review note
"I used AI to create a draft. Before using it, I checked the task, facts, source needs, assumptions, privacy risk, and final responsibility. Any claim that affects a real decision still needs a trusted source or a qualified reviewer."
Useful sources
- NIST AI Risk Management Framework for thinking about AI risk and trustworthiness.
- OECD AI Principles for human-centered and accountable AI use.
- FTC guidance on AI claims for avoiding unsupported claims about AI systems.