More than half of academic researchers now acknowledge using artificial intelligence tools when reviewing scholarly manuscripts, according to a recent global survey conducted by the academic publisher Frontiers. The poll, which gathered responses from roughly 1,600 researchers across 111 countries, highlights how deeply AI has begun to penetrate one of the most critical stages of scientific publishing: peer review.
The survey reveals that nearly one in four respondents have increased their reliance on AI-assisted tools for peer review within the past year alone. Published on December 11 by Frontiers, a Switzerland-based publisher headquartered in Lausanne, the findings confirm what many in academia have suspected as AI systems powered by large language models, such as ChatGPT, become increasingly widespread in research workflows.
Elena Vicario, director of research integrity at Frontiers, notes that the results underscore an important reality. Researchers are already integrating AI into peer-review tasks, despite widespread external guidance cautioning against uploading unpublished manuscripts into third-party AI tools. According to Vicario, acknowledging this behavior is a necessary first step toward managing it responsibly.
While some academic publishers, including Frontiers, permit limited AI use during peer review, they generally require reviewers to disclose such usage. Like most publishers, Frontiers prohibits reviewers from uploading unpublished manuscripts to public chatbot platforms. These restrictions are intended to protect confidentiality, safeguard sensitive data, and prevent potential violations of authors’ intellectual property.
The survey calls on publishers to adapt their policies to reflect what it describes as a “new reality” in scientific publishing. As AI adoption accelerates, the report argues that clearer, more practical guidelines are needed. In response, Frontiers has introduced an internal AI platform designed specifically for peer reviewers and made available across its journal portfolio. Vicario emphasizes that AI in peer review must be used with clear instructions, proper training, and well-defined human accountability.
Other publishers share similar views. A spokesperson for Wiley, based in Hoboken, New Jersey, stated that publishers should take a proactive role in communicating best practices, particularly around disclosure, to ensure transparency and responsible AI use. However, Wiley’s own survey earlier this year suggested that researchers still show relatively low interest and confidence in applying AI to peer review. According to the publisher, nothing in its current journal portfolio contradicts that assessment.
How researchers are using AI
The Frontiers survey also sheds light on how AI is being applied in peer review. Among respondents who reported using AI tools, 59 percent said they rely on them to help draft peer-review reports. Around 29 percent use AI to summarize manuscripts, identify gaps in arguments, or verify references. Another 28 percent turn to AI to detect potential research misconduct, including plagiarism and image manipulation.
Mohammad Hosseini, a researcher specializing in ethics and research integrity at Northwestern University Feinberg School of Medicine in Chicago, views the survey as a meaningful attempt to measure both the prevalence and acceptability of AI use in peer review. He notes that understanding how AI is being applied across different contexts is crucial for shaping effective policies.
Beyond surveys, some researchers are conducting hands-on experiments to test AI’s actual performance in peer review. In a recent example, engineering scientist Mim Rahimi from the University of Houston designed an experiment to evaluate whether the large language model GPT-5 could effectively review a paper he co-authored for Nature Communications.
Rahimi tested four different configurations, ranging from simple prompts asking the AI to review the paper with minimal context to more complex setups that included additional academic literature to help assess novelty and rigor. He then compared the AI-generated feedback with the real peer-review reports received from the journal and shared his observations publicly.
The results were mixed. While GPT-5 was able to replicate the formal structure and polished tone of a peer-review report, it struggled to provide meaningful, constructive criticism and frequently introduced factual inaccuracies. Notably, more complex prompting did not improve performance; in fact, the most elaborate setup produced the weakest review. Other research has similarly found that AI-generated reviews can resemble human ones at a surface level but often lack depth and detailed critique.
Rahimi concludes that while AI tools may offer limited informational support, relying on them alone for peer review could be harmful. Without careful oversight, such tools risk undermining the quality and reliability of scholarly evaluation.
A closed-system approach
Earlier this year, Frontiers began piloting its own internal AI tool for peer reviewers in select journals. Unlike public chatbots, this system operates within a closed environment designed to protect manuscript confidentiality and authors’ intellectual property. According to Vicario, the platform uses publisher-developed prompts to assist with specific tasks, such as summarizing submissions or assessing whether a paper aligns with a journal’s scope.
Use of the tool is optional for reviewers, and its role is deliberately limited. Reviewers are not permitted to copy and paste AI-generated text directly into their reports, as the system is not intended to replace human judgment or produce complete reviews.
Journal editors are able to see whether reviewers have used the AI system and can disclose this information to authors during the review process. Following its initial trial, Frontiers is now expanding the tool across all of its journals.
Rahimi views this approach as a potentially appropriate and timely response to the growing role of AI in peer review, though he remains curious about how effectively the system will perform at scale. Vicario acknowledges that the platform is still in its early stages and stresses that ongoing improvement and iteration will be essential as both AI capabilities and publishing practices continue to evolve.
Leave a Reply