To verify an AI output, check its important claims against evidence outside the generated answer. Open the cited sources, inspect calculations and compare the result with the original task. Fluent wording and a confident explanation do not establish accuracy.
NIST’s Generative AI Profile discusses false or erroneous generated content and recommends checking sources and citations. The workflow below is a practical starting point for reviewing business drafts; it does not guarantee that every error will be found.
Decide what would happen if the answer were wrong
A brainstorming list and a customer-facing product claim need different levels of review. Identify where the output will be used, who might rely on it and whether a mistake can be corrected before it causes harm.
For legal, medical, financial or security decisions, involve a person qualified for the task and use the required professional process. A second AI answer is not a substitute for that review.
Turn the draft into checkable claims
Highlight names, dates, prices, quantities, quotations, product capabilities and recommendations. Separate statements of fact from assumptions and suggestions. Identify omissions too: a plan can contain no obvious falsehood and still leave out a step that makes it unusable.
For example, an illustrative software comparison might say that a free plan includes an integration. That needs a current plan-specific source, not just a page showing that the product offers the integration somewhere.
Read the source behind each important claim
Open the source independently. Confirm that it exists, comes from the expected publisher and supports the exact statement. Check the date, relevant version, region and plan. A real citation can still be irrelevant or outdated.
Prefer the original documentation, research or record where available. Follow citations back to it instead of relying on several articles repeating the same unsupported assertion. If the evidence is unclear, narrow the claim, label the uncertainty or remove it.
Check the type of output
- Summaries: compare with the original material for missing qualifications, reversed meanings and invented details.
- Numbers: recompute from the source inputs. Check units, date ranges, totals and the denominator behind each percentage.
- Quotations: compare the words and attribution with the original source.
- Code: review it and test in an appropriate isolated environment before use. Passing one example does not establish that every path works.
- Instructions: check prerequisites, permissions and what happens when a step fails.
Use prompts to expose gaps
Ask the system to work from supplied material, identify missing inputs and separate sourced facts from assumptions. These instructions can make review easier, but the output still needs checking. Repeating “are you sure?” or asking another model to agree does not independently verify a claim.
Keep approved examples and known failure cases in a shared prompt library. Retest them when the tool or workflow changes.
Record the decision to use the result
For work that matters, save the reviewed version, important evidence, corrections and reviewer. Track recurring failures so you can change the inputs, task boundaries or approval process.
If a claim remains unsupported, do not publish it as established fact. Decide whether more research, specialist review or a narrower answer is needed. Use human approval before the downstream action when waiting until afterward would be too late.