By Carmino de Souza
Image: Mariola Grobelska by Unsplash.
In my busy academic life, I am always on “both sides of the fence” i.e., as a research paper author and as a reviewer of articles submitted by numerous scientific journals around the world requesting my assessment. It is a highly responsible job, as we are always committed to excellence and quality in scientific information and to improving the submitted article.
The current standard in the academic publishing community is that authors can use generative artificial intelligence (genAI) in preparing submissions (with some caveats; see this position statement from COPE, Committee on Publication Ethics, and publishers). However, there are rigorous restrictions on the use of AI in peer review.
Many journal policies specify that editors and reviewers should not present content from submissions to genAI tools, and some completely prohibit the use of genAI in peer review (e.g., Science) or allow specific cases of AI use, such as translating or editing the reviewers’ own comments.
Why this different standard? Policies that restrict or prohibit AI in peer review mitigate the risks that the use of genAI by editors or reviewers could introduce, such as:
- Breach of confidentiality of unpublished content and sensitive data;
- Loss of rigor and specificity in the evaluation process;
- Fraudulent representation of genAI results and peer review contributors; and
- Enabling and accelerating peer review manipulation (e.g., by “paper mills”)
It may seem paradoxical or even hypocritical that journals and publishers are exploring options for the internal use of AI in peer review. A key difference between internal use (by journal staff) and external use (by academic editors and reviewers) is that a journal can implement internal tools in a controlled technological environment that protects data security, so that confidential content is not entered into training sets, which would affect the results of other users.
When data security measures are implemented, AI can help improve the consistency with which journals apply their standards and policies. For example, AI can detect and produce review reports questioning issues such as incomplete, unverifiable, or retracted references, problematic statistical analyses, and non-compliance with data availability and pre-registration requirements.
While there are several compelling use cases for AI in supporting peer review, humans remain indispensable for providing rigorous content evaluation. While genAI detects and averages preexisting content, humans innovate and evaluate. We introduce new ideas and perspectives, bring creativity, curiosity, and intellectuality, and are able to synthesize, contextualize, interpret, and critique based on knowledge spanning multiple domains.
In essence, machines are far from being able to replicate human cognition, and therefore humans can engage in peer review and scientific discourse in ways that machines cannot. In practical terms, this means that people can identify issues that would not be apparent to a machine reader or algorithm, and that may be crucial to scientific validity and integrity.
That said, the transition to a hybrid model of peer review, with humans and AI, could mitigate known issues with peer review, including the heavy burden peer review places on scholars and longer-than-ideal review times. If AI covers the technical aspects of an evaluation, we may be able to use fewer reviewers to cover aspects of peer review that require uniquely human executive functioning capabilities.
As proof of concept for this model, a talk at the 2025 Peer Review Congress discussed a “Fast Track” peer review offering from NEJM AI, in which decisions are issued within a week of submission, based solely on the editors’ assessment of the manuscript and two AI-generated reviews. While one-week turnaround times are attractive, there are several reasons to include at least two human experts in peer review, either as editors and/or reviewers.
Authors and papers benefit from evaluations that reflect different (human) perspectives; often, multiple individuals are needed to cover the subject matter and methodological expertise required for a rigorous evaluation. It is important to note that having two or more humans involved in peer review also increases the likelihood that any important scientific and integrity issues will be identified and lends greater overall credibility to publications and journals. It also offers a degree of protection for authors, journals, and the community at large against issues that could compromise peer review, such as personal biases, conflicting interests, poor-quality reviews, and unethical (mis)use of peer review for personal gain.
The age of AI may be here to stay, and publishers and researchers will continue to explore its uses, but caution and careful consideration are needed at every stage of the peer review process. And ultimately, it will never replace a person’s knowledge and judgment.
Original article in Portuguese
Os perigos do uso da IA na revisão por pares
About Carmino Antonio De Souza
Carmino Antonio De Souza is a full professor at Unicamp. He was Health Secretary for the state of São Paulo in the 1990s (1993-1994) and for the city of Campinas between 2013 and 2020. He was Executive Secretary of the Special Secretariat for Science, Research, and Development in Health of the state Government of São Paulo in 2022 and current Chairman of the Board of Trustees of the Butantan Foundation. He is currently Scientific Director of the Brazilian Association of Hematology, Hemotherapy, and Cell Therapy (ABHH) and Principal investigator for CEPID-CancerThera, supported by FAPESP.
Translated from the original in Portuguese by Lilian Nassi-Calò.
Related Posts:
Como citar este post [ISO 690/2010]:
Recent Comments