.jpg){kind=link}
- Published
- Updated
An AI-generated monitoring summary can read well while adding unsupported conclusions or presenting an open issue as resolved. Human review is more than language editing. Compare the summary with its sources to confirm that facts, responsibilities and uncertainties important to decisions remain intact.
Scope of the decision
Define intended use and an authorised data environment in advance. Reviewers should know which notes and data cut were summarised. When sources conflict, a model’s choice of one fluent narrative is not an appropriate resolution. Remove, correct or explicitly qualify statements that cannot be verified.
Worked review
For example, a monitor’s note says that additional site evidence is pending, while the generated summary says the issue was resolved. The reviewer does more than edit that sentence: similar closure claims are examined in other outputs. Original records and authorised decisions remain authoritative rather than being replaced by generated text.
Practical checks
- Trace conclusion words such as complete, acceptable and no impact to their sources.
- Check accurate transfer of participant, site, date and action ownership.
- Test whether missing or conflicting information disappears from the summary.
Acceptance and limits
Accept a summary as supporting text only when it is traceable and appropriately reviewed. A human signature does not make an unsupported claim correct. Important errors should trigger reassessment of scope, related outputs and limits on use.
Frequently asked questions
- Is grammar review sufficient?
- No. Fluency does not establish factual accuracy or preserve the meaning of a decision.
- Is every sentence with a source link correct?
- No. The linked material must actually support the claim; a source name alone is insufficient.