This February, ecologist Timothe Poisotto was surprised when he read a peer review of the manuscript he submitted for publication. One of the judge's reports is written in artificial intelligence (AI), or perhaps fully written. It contained the clear statement, “Here is a revised version of the review with improved clarity and structure.” This strongly indicates that the text was generated by the leading language model (LLMS).
Poisot has not told journal editors his suspicions. He asked what the relevant journals (prohibiting the use of LLM in peer reviews) were not revealed in this article.
However, in a blog post about the incident, he strongly opposed automated peer reviews. “I submit my manuscript for reviews in the hopes of getting comments from my peers. If this assumption is not met, the entire social contract for peer reviews is gone,” writes Poisot, who works at the University of Montreal, Canada.
AI systems are already transforming peer review. This is encouraged by the publisher and sometimes violates the rules. Both publishers and researchers test AI products to flag errors in the text, data, code and reference of manuscripts, guide reviewers to more constructive feedback and hone their prose. Some new websites offer a one-click full review created by AI.
However, there are concerns with these innovations. Today's AI products are cast in the role of assistants, but AI will eventually become dominant in the peer review process, completely reducing or reducing the role of human reviewers. While some enthusiasts view peer review automation as inevitability, many researchers, such as Poisot and Journal Publishers, see it as a disaster.
My other editor is AI
Even before LLMS-based CHATGPT and other AI tools came into being, publishers used a variety of AI applications to facilitate the peer review process for over a decade, including tasks such as checking statistics, summarizing findings, and mitigating peer reviewer selection. However, the emergence of LLMS, which mimics fluent human writing, has changed the game.
In a survey of around 5,000 researchers, about 19% said they tried to “increase the speed and ease” of reviews using LLM. However, this study by Wylie, a publisher based in Hoboken, New Jersey, did not question the balance between using LLM to cover prose and relying on AI to generate reviews.
Three AI-powered steps to faster, smarter peer review
One study1 Peer review reports of papers submitted to AI conferences in 2023 and 2024 found that between 7% and 17% of these reports contained signs that they had been “significantly changed” by LLM.
Many funders and publishers are currently banning grants and paper reviewers from using AI. However, if researchers host offline LLMS on their own computers, the data is not fed into the cloud, says Sebastian Poldammann of the University of Copenhagen.
Paraphrase the notes using offline LLMS, as long as LLM “doesn't crank out the full review for you,” writes Dritjon Gruda, an organizational researcher at the Catholic University of Portugal, in Lisbon. Nature Career column.
But “taking superficial notes and synthesizing LLM is far below writing proper peer reviews,” rebuts Carl Bergstrom, an evolutionary biologist at Washington University in Seattle. There is a risk that reviewers can skip most of the process of relying on AI to create reviews, providing shallow analysis. “I'm thinking about writing,” says Bergstrom.
LLMS certainly can improve the style of some reviewers, Pordam Mann said. However, LLM output most often contains errors as the tool works by creating text that appears to be statistically likely based on training data and input.
The gap between humans and LLM is not particularly large, according to studies that have provided reviews of their papers to more than 300 US computational biologists and AI researchers in many cases.2. Approximately 40% of respondents said that AI is more useful or useful than human reviews. Additionally, 42% of people have AI that is less useful than many, but more useful than some (see Comparison of AI and Human Peer Review).
Source: ref. 2
AI that goes beyond editing
The team behind the study, led by James Elephant, a computational biologist at Stanford University, California, compares AI with Human Reviews, is currently developing a reviewer called “feedback agent.” It evaluates human review reports against a checklist of common issues such as ambiguous or inappropriate feedback, and then suggests how reviewers can improve their comments.
At the publisher's innovation fair held in London last December, many AI developers lined up with their products to improve peer reviews more than just editing. One tool called Eliza was launched last year by the company's World Brain Scholars (WBS) in Amsterdam, Netherlands, and makes suggestions for improving reviewer feedback, recommends relevant references, and converts reviews written in other languages into English. The tool is not intended to replace human peer reviewers, says Zeger Karssen, founder of WBS. “The tool simply analyzes what the peer reviewer writes down,” he says.
A similar tool is a review assistant developed by multinational publishing services companies Enago and Charlesworth. Initially, the tool used the LLM system to answer structured queries about the manuscript and was able to be reviewed or verified by reviewers. However, after talking to the publisher, the developer added the “human first” mode. In this mode, the reviewer answered the question before the AI tool looked at the answer. The tool “can support what is done in a legitimate way, in a legitimate way by reviewers who may already be doing it illegally,” says co-developer Mary Miskin, global operations director at Charlesworth, based in Huddersfield, UK.
Another AI approach aims to free reviewers from the hassle of peer review. A startup called Grounded AI in Stevenage, UK, has checked whether a paper cited in Manuscripts exists and used LLM to develop a tool called Authenticity that analyzes whether the cited works correspond to the author's claims. It works like “a workflow that a motivated, strict human fact checker experiences if it's always there in the world,” says co-founder Nick Morley.
What is the best AI tool for research? Natureguide
Also, many efforts have emerged to apply LLM support tools to existing papers, from software to spot image replication to statistical checking programs. However, researchers have expressed concern that LLM may not be reliable and that some obvious errors may be false positives.
One of the AI review tools that is already in trial with publishers is the alchemist review developed by a grounding AI and company called HUM in Charlottesville, Virginia. The software authors say it can summarise core findings and methods, assess the novelty of the research, and validate citations. They also say that reviewers can use the tools in a secure environment that protects the confidentiality of manuscripts and authors' intellectual property.
AIP Publishing, the publishing division of the American Institute of Physics, based in Melville, New York, pilots the software version in two journals, says Ann Michael's Chief Transformation Officer. Journal editors will test prototypes of the tool and, at their discretion, can be tested by some peer reviewers. However, publishers do not test the ability of the tool to judge novelty, as they suggest that editors do not value editors as useful as other features. “We are trying to learn how to responsibly apply AI to peer review,” she says. This tool emphasizes that it is used before human reviews.
Other publishers also spoke Nature That they were exploring the development of in-house AI tools for peer review, but didn't say exactly what they were working on. For example, Wiley is “examining a variety of potential use cases for AI to enhance peer reviews, including the editor level and reviewer level,” the spokesperson said.
Survey of guidelines in the Top Medical Journals in December 20243 Of the large publishers, Elsevier has now found that reviewers prohibit the use of generated AI or AI-assisted reviews, while Wiley and Springer nature allow for “limited use.” Both Springer Nature and Wiley need to disclose their use of AI and prohibit online uploading of manuscripts to support reviews. (NatureThe news team is editorially independent from publishers. ) In this study, 59% of the 78 top medical journals focused on peer review guidance on banning problems. The rest allow it and there are various requirements.
AIRED review?
The most fundamental application of AI in peer review is a tool that directly provides automated reviews of manuscripts. One example is the Paper Wizard, which generates a full-scale multi-page review when paper is uploaded, checking detailed aspects of the methodological design, such as statistical rigor. Shane Ehard, a co-creator cognitive neuroscientist in Brisbane, Australia, says it is a “pre-review” product aimed at authors supporting authors with their own work.
