.jpg){kind=link}
- Published
- Updated
A language model may fill missing safety information with plausible details. Such additions can be difficult to notice inside fluent text. Narrative review should check source support for dates, treatment, event course and outcomes while keeping missing information explicitly unresolved.
Scope of the decision
Assess automated drafting only within a defined and authorised use. A model does not own clinical judgement or reporting decisions. Medical plausibility is not enough to add a detail to a study record. Preserve appropriate clinical review and control over the resulting document.
Worked review
For example, the source states that the outcome is unknown, but the draft says the participant recovered. The reviewer removes that claim and keeps the outstanding follow-up visible. Medicine changes and dates elsewhere in the draft are also traced to source; correcting one wrong sentence does not establish that the entire narrative is verified.
Practical checks
- Verify which source statement supports every relevant clinical detail.
- Check that unknown or pending assessments have not become definite outcomes.
- Examine other drafts for similar behaviour after identifying a critical error.
Acceptance and limits
Acceptance requires preservation of source-supported meaning and completion of appropriate authorised review. A model confidence score is not source evidence. Use the proper follow-up process for missing information rather than allowing generated prose to conceal unexplained gaps.
Frequently asked questions
- May a medically plausible detail be added?
- Not as a study fact without source support; appropriate expert assessments follow their authorised process.
- Does a human signature remove every risk?
- No. Review scope, competence and source comparison must be substantively adequate.