Build an AI-assisted review governance protocol by writing three sections before any AI tool is used: a permitted-use list, a human verification requirement for each output, and an oversight log that records a source location for every AI-assisted claim.
- Define permitted and excluded uses. List the review tasks where AI may assist, such as drafting summaries, screening titles and abstracts, or checking reference metadata. Name the tasks where AI output cannot enter the manuscript without a named reviewer's sign-off. Tie each permitted use to the review stage it serves, so a reviewer can tell from the protocol whether a given action is in scope.
- Set the verification standard per output type. Match verification to the claim. For risk-of-bias judgements, the Cochrane RoB 2 approach requires signalling questions and written justifications that support a low risk, some concerns, or high risk result; a reviewer should confirm that the justification, not only the label, is present. For reporting transparency, PRISMA 2020 sets out the checklist and flow-diagram items covering why the review was done, what methods were used, and which studies were found and included; check AI-drafted methods and results text against those items. For reference lists, use a citation check such as WisPaper TrueCite, which checks BibTeX against real academic databases to flag references that need verification, then confirm each flagged entry at its source.
- Log every AI output with a source location. Use an oversight log template with these columns:
- Output ID and date
- Tool and version
- Task and prompt summary
- Manuscript location (section, paragraph, or line)
- Source location supporting the claim (page, section, table, or record identifier)
- Reviewer name and verification outcome
- Status: accepted, revised, or escalated
- Apply the escalation rule. If an AI-assisted claim has no traceable source location, mark it escalated and hold it out of the manuscript until a reviewer locates the supporting passage or removes the claim. Record the escalation in the log with the reason and the resolving action.
- Review the log at each draft milestone. At each manuscript or grant milestone, check that every AI-assisted entry has a reviewer name, a source location, and a closed status. Entries still open at that point become the agenda for the next reviewer meeting.