Peer review in the age of AI: the good, the bad and ugly

Author: Tim Kersjes

Head of Research Integrity, Resolutions

Observations on AI and peer review from the 10th International Congress on Peer Review and Scientific Publication 

Earlier this month I had the pleasure to attend the 10th Peer Review Congress in beautiful Chicago. The Peer Review Congress brings together editors, meta-researchers, publishers and technologists from across the world, and, like previous editions, this year was packed with excellent presentations on scholarly publishing. That inevitably meant that a big part of the program was dedicated to research integrity. And of course, like every conference in the last two years, AI was a major theme, similar to this year's Peer Review Week, Rethinking Peer Review in the AI Era. 

Several presentations at the Peer Review Congress were about studies on the use of AI by researchers or on how AI use was declared and acknowledged, and mirrored the feedback we have received as well.  A sizeable survey presented by the editorial team at The BMJ found that for authors who had disclosed the use of AI, 87% of them stated that they had used AI to improve the quality of writing, and a study by editors at JAMA reported something similar. Another similar survey involving Chinese medical researchers found that many researchers were hesitant to use AI or to self-report it on submission, noting a lack of consistent AI usage guidelines, with many different journals applying various levels of permitted usage and disclosure requirements. This hesitancy was echoed in a recent Nature survey of over 5,000 researchers, which found that although 65% considered it ethically acceptable to use AI to generate text, only 8% had done so to draft a manuscript—and most chose not to disclose it. These findings underscore a growing disconnect between ethical acceptance and actual practice, reinforcing the need for clearer, harmonised guidance across the publishing ecosystem. Publishers, including Springer Nature, are currently working together on precisely this through STM, who have released for feedback a draft report detailing a clear framework to help publishers define, evaluate, and guide the transparent use of AI in manuscript preparation. 

The same concerns apply to the use of AI in peer review. We’re already seeing some instances of AI being used in the peer review process in problematic ways: it can lead to sloppy and lazy peer review, with reviewers simply letting ChatGPT to do the work.   However, we shouldn’t think it’s not all be bad.  One relatively small-scale study by the BMJ indicates that peer review reports by LLMs 'matched or exceeded human reviewers on a few key dimensions of review quality’. These indicators included identifying strengths and weaknesses, as well as providing useful comments on the writing or the article and the organisation and presentation of the data.

Still, I believe that, fundamentally, peer review must ultimately remain a human endeavour. At its heart, peer review should be a fair and critical evaluation of a manuscript by the authors’ peers, with the ultimate goal to improve the quality of the article. Looking at it like that, simply putting papers into ChatGPT means something is lost. It’s not only about human accountability, there are other aspects that only humans can add:  

• Contextual understanding: AI may struggle to grasp the nuances of interdisciplinary research, emerging methodologies, or the significance of a study within its field. As the old adage goes, there’s no substitute for experience. 
• Ethical judgment: Machines can flag inconsistencies, but they cannot weigh the moral implications of a study’s design or its societal impact. 
• Empathy and tone: Constructive feedback is about guiding authors toward improvement as well as pointing out flaws. This requires emotional intelligence and cultural sensitivity. 

We shouldn't discount the numerous ways AI could support peer reviewers, as long as clear ethical boundaries are set and technologies used responsibly. This is something many of my colleagues are working on and some elements are already in testing.  

Integrity is at the core of how we view scholarly publishing. As AI itself is evolving, researchers’ and reviewers’ usage of and attitude towards the use of AI is also changing, and we need to reflect that, whilst centring our responsible AI policy and focus on integrity. This means that we continue to: 

• Evolve and communicate our policies on AI use in authorship and review  
• Support researchers with training and guidance on responsible AI use  
• Invest in tools that detect manipulation without compromising fairness 

And throughout this ongoing time of change, we must ensure that human insight, compassion, and accountability remain at the heart of the process.