What Is the Role of AI in Peer Review? — Companion Materials

The reviewer reports, study design, and reanalysis record behind the note

These are the working materials behind the note What Is the Role of AI in Peer Review?. They are provided so the comparison can be inspected and, where the underlying data are open, reproduced. The study is a single worked example (n = 1), not a validation of AI review.

This page collects the artifacts the note refers to: the four AI reviewer reports, the de-identified human reviews used as study inputs, the protocol and run package that produced the AI panel, and the record behind the independent reanalysis. Each file is offered as a download. Third-party copyrighted material — the cited-reference PDFs, the submitted manuscript proof, and the trial’s raw-data workbook — is not rehosted here; where it is public, it is linked at its source.

The target paper is Zhang et al., Prophylactic tranexamic acid for the prevention of postpartum haemorrhage in women with placenta praevia (BMJ 2026;393, doi:10.1136/bmj-2026-089636). The BMJ practices open peer review, so the signed human reports, decision letters, author replies, and both manuscript versions are published alongside the paper.

The AI reviewer reports

Four reviewers, each a separate, fresh instance of GPT-5.5 (gpt-5.5-2026-04-23, reasoning effort high), run fully offline with no web access. Each saw only what the human reviewers had at submission — the submitted manuscript proof with its bound statistical appendix, plus the full text of the cited references (22 of 24 available; references 5 and 16 were marked uncheckable). The files below are the complete transcripts, including the harness’s developer message and the model’s document-retrieval tool calls, so the conditions of each review are fully visible.

Role Report
Trial statistician ai-review-statistician.md
Obstetrics / maternal-health methodologist ai-review-obstetrics.md
Generalist clinical reviewer ai-review-clinical.md
Literature / citation reviewer ai-review-citations.md

The human reviews

The signed original reports are published in The BMJ’s open peer-review record. The versions below are the de-identified copies used as inputs to the blinded comparison: reviewer names, institutions, competing-interest blocks, and journal boilerplate were removed so the analyst could not tell which panel was which. Substantive scientific and process comments are unchanged.

Study design and run package

How the study was designed and the AI panel run, written before the panel was assembled. Each reviewer’s prompt was written from its role, not from the points the human reviewers actually raised, to avoid manufacturing a match. The full prompt for each reviewer is reproduced verbatim at the top of its report.

  • study-design.md — the thesis, design principles, the blind Step 1 panel, and the seven-step analysis pipeline
  • run-log.md — the operator’s harness-and-run log (model, parameters, what each instance could see)

The reanalysis

The note’s independent check of the analysis was run against the trial’s raw data, which The BMJ released with the published paper (see the paper’s data-availability statement). The workbook is the journal’s file and is not rehosted here. The record below documents that the workbook reproduces the paper’s headline numbers exactly and fixes the group coding used in the reanalysis.