A citation is a place to start checking, not a reason to stop

Check what the source actually supports

Record the exact claim before you open the source. Then find the passage, table or chart that is supposed to support it. If you cannot locate support, mark the claim unresolved. An inaccessible page is not proof the claim is false, but it is not evidence you have verified either.

OpenAI's prompting documentation describes providing relevant context and constraining responses to a chosen set of resources. That can help define a source-checking task. It does not mean the output has passed a factual check. In this lesson, the final judgment remains a comparison you make against the source.

Keep the population and the time period attached

Here is a fictional example, not a real survey: a report says 30 of 100 people attending an optional workshop completed an exercise. The statement Thirty percent of all employees completed the exercise does not follow. The source only describes workshop attendees; it gives no count for the whole workforce.

The arithmetic is not the problem. Thirty divided by one hundred is thirty percent. The unsupported step is replacing workshop attendees with all employees. Also check whether the source describes a day, a month or a different period from the claim. A correct number can still describe the wrong thing.

Keep reporting separate from demonstration

A company announcement can establish what the company says. A product demonstration can show what happened in that demonstration. Neither automatically establishes how the product performs for every user. Write the narrower statement when that is all the evidence supports.

Use supported, contradicted or unresolved as your working labels. Save a short explanation beside the label. Asking another AI model to agree can be useful criticism, but agreement is not a replacement for the original evidence.

Make your next summary easier to inspect

When using AI for a sourced summary, ask for the claim, source URL and supporting passage separately. Open the links yourself and check important quotes exactly. If the source changes or cannot be accessed, preserve that limitation rather than silently filling the gap.

For recurring work, keep a few examples of supported and unsupported claims and test your process against them. OpenAI's evaluation guidance recommends task-specific evaluation and human checks. A small checklist you actually use is more useful here than a score with no explanation.

Find the unsupported leap

Use the fictional example first. Do not paste confidential work into an unapproved tool.

Compare this claim with the fictional source excerpt. Classify it as supported, contradicted or unresolved. Explain any mismatch in population, time period or quantity. Do not add outside facts.

SOURCE: Of 100 people attending an optional workshop, 30 completed the exercise.
CLAIM: Thirty percent of all employees completed the exercise.

UNRESOLVED. The excerpt supports 30 percent of workshop attendees. It does not give the total workforce or completion count for employees who did not attend. The claim about all employees cannot be established from this excerpt.