How Cochrane’s Rapid Reviews Methods Group Sets Boundaries for AI Use

The clearest current role for AI is quality assurance. The authors point specifically to single-reviewer workflows, standard practice in many rapid reviews because of time constraints, as the setting where AI adds the most defensible value. Cochrane’s own rapid review guidance already permits switching to single-reviewer screening once inter-rater agreement during dual screening is high, a shortcut that will predictably miss some eligible studies. In that context, AI functioning as a check on a single reviewer, flagging studies that may have been missed, surfacing inconsistencies, catching likely extraction errors, is treated as a mitigation for a known and accepted risk rather than a new capability being added for its own sake.

The evidence on AI performance does not yet support broader use. The statement is explicit that published evaluations of generative LLMs across search strategy development, screening, risk of bias assessment, and data extraction show highly variable accuracy, ranging from strong performance on some tasks to concerning error rates on others. That inconsistency is a central reason the RRMG stops short of recommending AI for tasks beyond quality assurance.

Any AI use has to be disclosed and justified. Review teams are expected to name, in the protocol, which tools they are using, what task each one performs, and how a human verifies the output. Cochrane rapid reviews are expected to include a dedicated AI Use Disclosure section, and if a generative LLM is involved, the model version and the prompts used need to be recorded.

Authors remain fully accountable. The statement is direct on this point: AI is a support tool, and responsibility for the accuracy and integrity of the review sits with the author team regardless of which tools were used or how well they performed.

The tool has to be shown to preserve rigor, not just save time. Teams choosing to use an AI tool are expected to be able to demonstrate that it upholds the review’s methodological integrity as a condition of using it at all.

The RRMG statement complements a broader position issued by Cochrane and other evidence synthesis organizations on AI use across evidence synthesis generally, and it points authors toward the RAISE framework for a more detailed treatment of acceptable AI use at each review stage. The authors also frame this as an interim position. The RRMG plans to revisit and update its guidance at least every two years as performance data on these tools matures.

For teams conducting rapid reviews, the practical implications are concrete:

• Tool disclosure. Name the tools being used in the protocol.
• Task specification. State what task each tool performs.
• Human verification. Describe how a human verifies the tool’s output before it is used in the final report.
• Model and prompt documentation. If a generative LLM is involved, record the model version and the prompts applied.
• Provisional status. Treat any AI-assisted output as provisional until a reviewer has checked it, not as a finished result.

The underlying standard the RRMG is applying is not new. It is the same standard that governs any other methodological choice in a rapid review: whether the decision can be explained, defended, and traced back to a person who is accountable for it. AI is useful in proportion to how well its outputs are verified, and the statement treats it accordingly, as a new instrument subject to the same scrutiny as any other.