Peer review needs a revolution. AI is already driving it

Challenges to peer review affect the quality and efficiency of academic publishing. AI tools already help reviewers and publishers but are we prepared for the future?

Photo by Pablo Zuchero on Unsplash

Imagine an article about peer review that didn’t follow the usual tropes. One that didn’t open with the quote that like democracy it is imperfect, but the best system we have. It wouldn’t close summarising that until a better system is found, we would do well to understand its strengths and weaknesses.

I’m confident that this audience is aware of the flaws of peer review, and the many challenges it has faced since its broad adoption in the 1960s. My focus here will be current issues, and how AI could change the picture. Unsurprisingly, I conclude that it could both help and hinder and governance and transparency are critical.

The challenges

Research integrity and researcher overload are two current, highly discussed issues in peer review and AI is making inroads in both areas. This post will look at AI supporting publishers, finders and academics, rather than entirely replacing the human. Jon Treadway wrote about that back in October.

Overworked academics are asked to do more and more and time is in short supply. Researchers report that they get more requests for peer review now than they used to, and many would like their work in this area to be better acknowledged. In a Springer Nature survey of 6,000 scientists, 70% wanted their peer review contributions to be considered in their evaluations, but only half report that this is the case.

While some publishers have offered discounts for future APCs to reviewers, or offer awards for peer review, others have begun to use ‘structured peer review’ to focus reviewers on the right questions and improve efficiency. Debates about financial compensation for researchers’ time continue, but given the volume of work undertaken, it seems unlikely this would be an affordable route for publishers, and funders would struggle to justify this use of resources for a service that has previously been ‘free’.

Improving efficiency

More realistic in ambition is the efficiency that AI could bring in supporting reviewing work. Tools are being developed to do initial checks for plagiarism, statistical errors, image manipulation, and whether submissions match journal scope before human reviewers see them, which could reduce the burden on human reviewers. The STM Integrity Hub lists tools that help safeguard the integrity of science. Several incorporate AI and support peer review; for example, Manuscript Manager uses AI tools for technical checks and reviewer identification; Clear Skies Papermill Alarm will scan submissions to alert to content that is similar to other papermill content; and, Springer Nature announced two tools that check for text consistency and image duplication.

OpenRxiv, which runs the bioRxiv and medRxiv preprint servers, recently integrated an AI-driven tool to give quick feedback on submitted papers – usually within 30 minutes. Developed by the start-up company q.e.d Science, the tool assesses originality, identifies gaps and suggests both experimental additions and edits to the text.

For reviewers, the adoption of AI is well established and ahead of regulation. A December 2025 report by Frontiers, Unlocking AI’s untapped potential: responsible innovation in research and publishing, argues that the potential of AI to support peer review is untapped and with responsible integration it offers a great opportunity. Over half, 53%, of respondents to the survey behind the report use AI in reviewing, and, of these, 70% for writing and only 25% for analysis, design or methodology. When asked about barriers to using AI in review, a lack of guidance or unclear rules were the most commonly cited reasons.

Undermining integrity

Academics can wait months for journal peer review to be completed. Yet despite apparent widespread AI adoption and the promise of improved efficiency, they are perturbed. The Institute of Physics (IOPP) report, AI and Peer Review 2025 – Insights from the global reviewer community, is also based on survey responses. Of 348 respondents, 57% would be unhappy if a reviewer had used an AI to review a paper they had contributed to and 42% unhappy if it were used to augment a reviewer report. Opinion on the impact of AI on peer review was divided: 41% felt it had a positive impact and this is an increase from IOPP’s 2024 survey, 37% a negative one and 22% were unsure.

Elsewhere there are growing concerns that AI use by reviewers impacts integrity. Last month Nature Index reported a study suggesting it is nigh on impossible to tell if a review is AI generated. Scientists fed the openly available reviews for 20 of eLife’s papers into Claude, Anthropic’s LLM, and then asked it to generate reviews of its own. When compared with the original reviews, they looked plausible but had no depth. Worryingly, checks with ZeroGPT and GPTzero, tools that aim to spot use of LLMs, failed to spot they had not been written by a human. ZeroGPT found 60% had been written by a human and GPTzero 80%.

It may be true that some of the human and AI errors in review are similar – both can be biased, and quality will vary, AI hallucinations and made up references or citations will undermine trust.

Governance and transparency

Peer review is changing. It has been challenged for as long as I can remember, yet no one can suggest a better way. Perhaps AI offers the change we need, although we have been caught unprepared and risk remaining that way. Current use is largely ungoverned and intransparent, and the technology moves so fast it can be hard to keep up. So how will we manage what is around the corner? Without oversight, AI could speed up review without improving quality, which opens a whole new can of worms for an already strained system.

If peer review is to be fair and accountable, strong governance is needed, and fast. Researchers should know if AI is evaluating their work and the role it plays in decisions. Governance should address: who is responsible for AI errors; how authors can challenge AI-based decisions; and, what level of explanation must systems provide for their assessments? Publishers and researchers must not leak unpublished information via AI tools and that submitted work isn’t improperly used to train commercial AI systems.

Most importantly, it has often been said that peer review is a system based on trust. Without transparency in the above, this trust will be greatly diminished.