Evidence & judgement
A genuine citation can be attached to an unsupported claim
Opening a source is the beginning of verification. Check whether its population, measure and conclusion match the sentence you want to publish.
A citation can exist, open correctly and still fail to support the sentence beside it. When you review an AI-assisted research summary, finding the paper is only the first check. You also need to compare what the author measured with what the summary says happened.
NIST’s generative AI profile identifies confidently presented false content as a risk and warns that generated citations or explanations can reinforce misplaced trust. Its action MS-2.5-003 recommends reviewing sources and citations in AI outputs during testing and monitoring.
Read the sentence before the bibliography
Start with the precise claim that matters to your decision. Mark the people or cases it concerns, the outcome and any causal language. Then find the source passage that supports each part. A title containing the right keywords is insufficient.
Consider a hypothetical summary: “AI training improves employee performance.” The linked paper might measure how confident students feel after a lesson. Confidence, learning and performance at work are different outcomes. Even if the reported confidence finding is sound, it cannot carry that broader sentence.
The repair is to narrow the claim to the measured result, identify the group studied and explain what remains unknown. If the result does not answer the business question, keep that question open rather than stretching the paper to fit it.
Follow the result back to its comparison
Read the methods and results together. Did everyone receive the same test? Was assistance available during practice, assessment or both? Was the comparison randomised, observational or a demonstration without a comparison group? These details determine which explanation the evidence can support.
Check what a reported percentage counts. A finding among participants who completed an exercise may leave out those who dropped out. Note whether the summary silently extends that result to everyone who started. Write the denominator and the stage of work beside the number while reviewing, even if the finished article can explain them more naturally.
Keep the evidence beside the approved wording
For a consequential claim, save the exact source version, relevant page or table, checked date and the wording you approved. Record whether you read the full text or only an abstract. If you cannot inspect the necessary result, describe that limit and withhold the unsupported inference.
Use that comparison when approving the sentence. For the separate problem of checking promises inside an operational draft, see the earlier guide to claims, omissions and authority.
Sources and further reading
Sources checked 3 November 2026. Prepared with AI assistance and editorial review.