AI Undermines Peer Review, Adds Confusion

Earlier this year I received comments on an academic manuscript of mine as part of the usual peer review process, and noticed something strange.

My research focuses on ensuring trustworthy evidence is used to inform policy, practice and decision making. I often collaborate with groups like the World Health Organization to conduct systematic reviews to inform clinical and public health guidelines or policy. The paper I had submitted for peer review was about systematic review conduct.

What I noticed raised my concerns about the growing role artificial intelligence (AI) is playing in the scientific process.

A service to the community

Peer review is fundamental to academic publishing, ensuring research is rigorously critiqued prior to publication and dissemination. In this process researchers submit their work to a journal where editors invite expert peers to provide feedback. This benefits all involved.

For peer reviewers, it is favourably considered when applying for funding or promotion as it is seen as a service to the community. For researchers, it challenges them to refine their methodologies, clarify their arguments, and address weaknesses to prove their work is publication worthy. For the public, peer review ensures that the findings of research are trustworthy.

Even at first glance the comments I received on my manuscript in January this year seemed odd.

First, the tone was far too uniform and generic. There was also an unexpected lack of nuance, depth or personality. And the reviewer had provided no page or line numbers and no specific examples of what needed to be improved to guide my revisions.

For example, they suggested I "remove redundant explanations". However, they didn't indicate which explanations were redundant, or even where they occurred in the manuscript.

They also suggested I order my reference list in a bizarre manner which disregarded the journal requirements and followed no format that I have seen replicated in a scientific journal. They provided comments pertaining to subheadings that didn't exist.

And although the journal required no "discussion" section, the peer reviewer had provided the following suggestion to improve my non-existent discussion: "Addressing future directions for further refinement of [the content of the paper] would enhance the paper's forward-looking perspective".

Testing my suspicions

To test my suspicions the review was, at least in part, written by AI, I uploaded my own manuscript to three AI models - ChatGPT-4o, Gemini 1.5Pro and DeepSeek-V3. I then compared comments from the peer review with the models' output.

For example, the comment from the peer reviewer regarding the abstract read:

Briefly address the broader implications of [main output of paper] for systematic review outcomes to emphasise its importance.

The output from ChatGPT-4o regarding the abstract read:

Conclude with a sentence summarising the broader implications or potential impact [main output of paper] on systematic reviews or evidence-based practice.

The comment from the peer reviewer regarding the methods read:

Methodological transparency is commendable, with detailed documentation of the [process we undertook] and the rationale behind changes. Alignment with [gold standard] reporting requirements is a strong point, ensuring compatibility with current best practices.

The output from ChatGPT-4o regarding the methods read:

Clearly describes the process of [process we undertook], ensuring transparency in methodology. Emphasises the alignment of the tool with [gold standard] guidelines, reinforcing methodological rigour.

But the biggest red flag was the difference between the peer-reviewer's feedback and the feedback of the associate editor of the journal I had submitted my manuscript to. Where the associate editor's feedback was clear, instructive and helpful, the peer reviewer's feedback was vague, confusing, and did nothing to improve my work.

I expressed my concerns directly to the editor-in-chief. To their credit, I was met with immediate thanks for flagging the issues and for documenting my investigation - which, they said, was "concerning and revealing".

Careful oversight is needed

I do not have definitive proof the peer review of my manuscript was AI-generated. But the similarities between the comments left by the peer reviewer, and the output from the AI models was striking.

AI models make research faster, easier and more accessible . However, their implementation as a tool to assist in peer review requires careful oversight, with current guidance on AI use in peer review being mixed , and its effectiveness unclear .

If AI models are to be used in peer review, authors have the right to be informed and given the option to opt out. Reviewers also need to disclose the use of AI in their review. However, the enforcement of this remains an issue and needs to fall to the journals and editors to ensure peer reviewers who use AI models inappropriately are flagged.

I submitted my research for "expert" review by my peers in the field, yet received AI-generated feedback that ultimately failed to improve my work. Had I accepted these comments without question - and if the associate editor had not provided such exemplary feedback - there is every chance this could have gone unnoticed.

My work may have been accepted for publication without being properly scrutinised, disseminated into the public as "fact" corroborated by my peers, despite my peers not actually reviewing this work themselves.

Timothy Hugh Barker is the chair of the RIPPER (Research Integrity and Predatory Practices in Evidence Reviews) Working Group. This group conduct research into the imapct that fraudulent and erroneous data and predatory journals have on evidence syntheses. He is the the deputy-director of the Adelaide GRADE Centre and a Senior Research Fellow of Health Evidence Synthesis, Recommendations and Impact (HESRI) at the University of Adelaide.