Assessment and decision
These activities have different meanings and must be identified separately in each paper's record.
Reporting and technical checks
Checks for RIMP scope, manuscript and disclosure rules, required information, readable files and declared permissions. They can flag missing information; they cannot establish that a scientific claim is true.
AI review
A disclosed model-assisted examination tied to evidence from the manuscript. It should identify uncertainty and unsupported claims, and record the tools and version used. It can be wrong and is not independent human peer review.
Human publication approval
A human's explicit decision to release the exact final package. This is an accountable publication decision; it does not turn the AI review into independent expert review.
A record for each version
A paper record identifies the version, publication status, assessment undertaken, evidence or report available to readers, and any correction or concern. An assessment of one version must not silently become an assessment of a later revision.
“AI assessed” is not a quality certificate. Multiple models may share errors or repeat unsupported conclusions. Their outputs must not be counted as independent human reviews or averaged into an automatic publication decision.
Decisions and disagreements
AI may prepare checks, identify questions and help organise a response. Final publication approval is by a human. Missing permission, unresolved ethical concerns, misleading claims or a version mismatch should be resolved before any release.
An author must be able to identify an error in an assessment and seek a correction. AI must not make a final misconduct finding or invent a reviewer. A human handles contested publication decisions and corrections.
Conflicts and independence
A relevant relationship can affect a review even when the reviewer acts in good faith. Conflict checks cover authorship, collaboration, supervisory relationships, personal or financial interests and other circumstances relevant to the paper.
A conflicted final decision maker must not approve the affected work. A separate authorised human decision is required before release; another AI does not remove the conflict. Read the conflicts policy.