Editors’ note: Today’s post is by Elena Vicario, Director of Research Integrity at Frontiers. Reviewer credit to Chef Haseeb Irfanullah.
Recent discussions on papermills and industrialized fraud have made one point undeniable: AI has become a powerful accelerant of research misconduct at scale. That reality is uncomfortable, but it is one the publishing sector needs to face directly. Investigations by Nature and Science into industrialized fake paper production have highlighted the scale and sophistication of the problem.
Organized fraud at this scale and sophistication demands a response that is equally systematic, and that includes the responsible use of AI. Basically, if AI is part of the problem, it also has to be part of the answer. During the past decade in research integrity, I have seen the use of AI in misconduct rise sharply. At the same time, AI is being used constructively by scientists to produce research and by publishers to prevent fake science from entering the scientific record.
Publishers increasingly rely on AI-supported systems to identify image manipulation, plagiarism, paper mill indicators, citation anomalies, and other integrity risks before publication. The STM’s Integrity Hub is one example of how AI-supported integrity screening is already becoming embedded across publishing workflows. From a research integrity perspective, the question is no longer whether AI belongs in scholarly publishing. It already does. The question is how it should be governed.
Yet, one part of the workflow remains especially contested: peer review. That caution is understandable. Peer review involves confidential, unpublished material, and concerns about data security, misuse, and over-reliance on automated tools are real. But it is increasingly difficult to defend blanket restrictions on reviewer use of AI when AI is already embedded elsewhere in the publishing process. My question is simple: why block conscientious reviewers from responsible AI use when, on either side of them, AI is already thriving and becoming indispensable to the publishing cycle?
The problem is that policy is lagging behind how people actually work. Reviewers are already using AI tools quietly to support parts of their assessments, whether to summarize manuscripts, sense-check statistics, interrogate code, or improve language and structure. Rather than denying that reality, everyone needs clarity about what is acceptable, under what conditions, and with what safeguards. Blocking AI use on misconduct grounds is not just inconsistent; it is counterproductive. The same system that relies on AI to detect image manipulation, papermill patterns, plagiarism, and statistical anomalies at scale, then turns around and tells reviewers not to use similar tools. That does not remove AI from peer review. It simply pushes it under the radar, unregulated. And that is where the real risk sits.
Peer review is the backbone of the scientific record, but it is under strain. Submission volumes continue to rise, and reviewer capacity struggles to keep pace while expectations of scrutiny increase. IOP Publishing’s State of Peer Review 2024, drawing on responses from over 3,000 researchers, found growing reviewer fatigue and workload imbalance as persistent structural problems.
Reviewers need the right tools. Used properly, AI is invaluable in the reviewer toolkit. Used invisibly or inconsistently, it creates new fault lines.
This is not, at heart, a technology problem. It is a governance and coordination problem. Funders, universities, and publishers are all responding to AI, but in different ways and at different speeds. Some permit limited use, others prohibit it, and many stay silent or remain unclear, leaving reviewers to navigate the problem alone.
Meanwhile, experimentation with AI-supported review workflows is gathering pace, including proposals for AI-reviewed preprints and automated evaluation models in open research. Fragmented use and policy do not slow AI adoption. They just make it less consistent and less transparent.
This is where organizations like the Committee on Publication Ethics can play an important role. COPE has long helped define the ethical baseline for scholarly publishing. It has already begun engaging with AI through discussion documents and position statements relating to generative AI and publishing workflows. But a gap remains in providing reviewers themselves with clear, operational guidance on what they can and cannot do with AI. The sector is still operating in a grey zone. What is needed now is not simply more acknowledgement that AI exists, but clear, practical guidance that helps editors, reviewers, and publishers work to a shared standard.
That means moving beyond the blunt “allowed versus prohibited” framing, which no longer reflects reality. Instead, COPE and others should lead on a set of shared expectations that people can actually work with:
- Transparency: meaningful disclosure of AI use, without penalizing responsible behavior or driving it underground
- Accountability: human reviewers remain fully responsible for their judgments and peer review reports, regardless of which tools supported the work. Journals should make this explicit in reviewer agreements.
- Boundaries: clarity on where AI can support (summarization, language, structure) and where it cannot replace judgement on novelty, significance, or scientific rigor
- Confidentiality and security: unpublished manuscripts should not be uploaded to tools that do not meet appropriate privacy and data-handling standards
- Oversight: editorial processes that make AI use visible, reviewable and auditable, just as other declarations and conflicts are already managed within publishing workflows
None of this is theoretical. Reviewers and editors are already navigating these questions in real time, often without clear guidance. That is not a sustainable position for a system that depends on trust.
Wheels are in motion, including at COPE and across the publishing sector, with STM consultation about AI use disclosure discussed at the World Conference on Research Integrity (WCRI) 2026. This is welcome, but the sector now needs leadership capable of translating discussion into shared standards and operational policy. COPE is well placed to convene that process from an ethical and governance perspective, while STM, publishers, institutions and funders need to help define what practical implementation looks like across real publishing workflows.
The current consultations and discussions at WCRI provide an important opportunity to begin establishing consensus around the core principles already emerging across the sector: transparency, accountability, confidentiality, oversight, and clear boundaries around where AI can and cannot support peer review. Those discussions should not end as conference talking points. They should become the basis for practical guidance and harmonized expectations that journals, editors and reviewers can effectively apply.
If the sector does not act, the gap between policy and practice will continue to widen. AI use in peer review will not disappear; it will simply remain uneven, insufficiently visible, and governed by inconsistent rules. That is a risk to trust.
The task is not to treat AI as either a solution or a threat to peer review. It is to define where it can appropriately support peer review, where it must be constrained, and how human accountability is preserved throughout.
What is needed now is timely, practical guidance that reflects how peer review is actually evolving and helps the sector respond with consistency and confidence.
Discussion
11 Thoughts on "Guest Post — Now is the Time for AI in Peer Review, and Publishing Policies Need to Recognize This"
It would be helpful to see the tools and checks developed by Dr. Vicario for Frontiers to see how these very broad ideas are implemented. We are facing similar issues as editors. Researchers can use AI in their process if disclosed. Publishers use it to detect and manage quality control. As editors we are prohibited from using it – even for checks of formatting and references and even in systems where data is protected and even to identify reviewers with specific characteristics beyond keyword overlap.
The article makes many important points. I would add that AI in peer review goes beyond being an arm’s length tool for independent use by volunteer peer reviewers.
If AI is truly a beneficial tool for conducting peer review (not just integrity checks etc.) it should be deployed in manuscript workflow *before* drawing on the scarce time and expertise of volunteer peer reviewers. Otherwise, journals are (a) just asking volunteers to do more work, not less, (b) exacerbating governance challenges, (c) creating duplicate work, and (d) abandoning the opportunity to define and shape a key function they are responsible for.
It is well known that transformative technologies can change underlying workflow and don’t just “plug replace” how we used to do things.
With all due respect, was this piece written by Dr. Vicario or by one of the LLM tools she thinks so highly of???
I actually had the same thought. Not to knock the author, but there are some sentences and verbiage that reads AI-ish to me. In a sense, this does reinforce her argument. It is so difficult to detect. I lean on the skeptical side of AI or specifically LLM usage, but the point is valid. AI is being used for peer review, but it remains under the radar because it is such a hot-button issue.
As an editor here, it is fascinating to me how often we’re now receiving questions/accusations about the provenance of posts, with assumptions that they must be AI-written. With our internal editors having seen this post go through multiple rounds of revision, it is clear, at least to us behind the scenes, that it is not (at least no 100%) AI-generated. But this strikes at the heart of an increasing problem in trust that AI is wreaking across both the scientific literature and the written word in general. I keep reading reports that kids these days are using the phrase “that’s AI” for anything they don’t like, and I’m concerned that this is increasingly becoming the case everywhere.
That said, we do not have any prohibitions on the use of AI at The Scholarly Kitchen, other than for images, but we do require transparency, and you may have noted addenda at the end of several recent posts detailing how AI was used in the writing process.
Very timely and important points raised here by Elena. The honest reality might be that AI is already used by reviewers in peer review and it may be the case of it not being disclosed adequately.
In my view, AI should be used to augment and not substitute. The value clearly lies in triage process, where methodological screening, integrity checks before review is done by algorithms and humans review novelty and far reaching aspects of these research that often comes with years of research in the field.
If there is one clear policy, then it should be disclosure of AI assistance, with humans in the loop accountable for judgement, and tools that discloses their reasoning transparently enough to be understood. In the interest of transparency, I work in this area (VerifyScience), so I would hold some view, but I will support the same argument irrespective of the tools used.
I am sorry I don’t agree. If AI is the reviewer, and AI writes papers, and probably AI reads papers, what do we need people for? Maybe just to write the AI programmes, but the AI will do that soon too. I’d argue to go the other way, and proscribe AI use for anything intellectually substantive. See: https://onlinelibrary.wiley.com/doi/10.1111/isj.70055 – no paywall, free to access – for an editorial on research integrity that very much gets to grips with this.
On the point of confidentiality and security, do any of the current tools have appropriate privacy and data-handling standards? If so, which ones?
The reality is AI probably writes many papers if not most (to some extent), AI probably reads most of papers (for summarising) and AI probably is used by most publishing houses for years to help their day to day task without formal acknowledgement.
It is just time to acknowledge and find ways on how it can be used to enhance day to day research by Humans who can spend more time asnwering more substantial and pressing questions of our times that has been stalled for years and moving fairly slowly since half a century despite better access to knowledge base both widely and in depth.
Without formal acknowledgement we set dangeroys precedent of having AI operating in shadows and reviewers shpoting in dark and retractions increasing undermining genuine substantive research.
It is not man versus machine argument anymore, it is man with machine argument at the moment.
Nice work. Journals should incorporate the new ghost detection framework, spectrum of AI use, and humanity test in their regular manuscript assessment work flow. Please read here: https://scholarlykitchen.sspnet.org/2026/01/22/guest-post-the-ghost-in-the-machine-why-generative-ai-is-a-crisis-of-authorship-not-just-a-tool/
A lot of the most valuable uses of AI happen before a paper ever reaches a reviewer to support a system under strain. As Elena and others have highlighted, publishers are already using AI-supported tools to identify plagiarism, image manipulation, paper mill indicators and other integrity risks. But most reviewers don’t know that, so it’s important to be transparent about the checks that have taken place so reviewers are clearer about where their expertise is most needed and what is asked of them. They don’t need to look for plagiarism or check journal scope. We need them to answer questions only subject experts can: How does this work sit within the wider literature? Does the methodology support the conclusions? What are the limitations? Where should the research go next?
I’d like to see a future where reviewers are provided with a secure AI-supported report alongside the manuscript and can build on it with their own assessment. That could reduce the time and effort required for each review while focusing attention on the areas where expert human judgement adds the greatest value.
There is absolutely a place for AI in peer review. The challenge is ensuring it is used safely, transparently and in ways that genuinely benefit reviewers, editors and authors. For me, the next step is not just better policy, but practical investment in AI-enabled tools that support reviewers while preserving human accountability.
