Artificial intelligence tools in scholarly publishing: guidance for peer reviewers

An Academic Medicine Podcast episode featuring this article is available from Apple Podcasts (https://podcasts.apple.com/us/podcast/academic-medicine-podcast/id1112697692) or wherever you get your podcasts. Peer review is fundamental to ensuring research excellence and integrity in scholarly publishing and serves as a mechanism for knowledge building in academic medicine.1,2 High-quality peer review brings relevant expertise to the evaluation of scholarship and has been demonstrated to improve manuscripts under consideration, benefitting both authors and editors.2,3 Peer review also acts as a crucial safeguard for rigor and helps minimize bias in the published literature.4 Moreover, engaging in peer review nurtures the professional development of reviewers themselves, sharpening critical appraisal skills and deepening their understanding of current and emerging conversations in the field.1,2 Peer review is, nevertheless, a time-intensive professional activity, and peer reviewers themselves have increasingly become a limited resource.5,6 The proliferation of journals, steady rise in manuscript submissions, and increasing demand for rapid decision-making and publication have placed an unprecedented strain on peer review. Concerns related to an inadequate number of expert peer reviewers, the quality and brevity of reviews, and reviewer fatigue are now common.7 These challenges place a burden on peer-review processes and threaten their role and sustainability in scholarly publishing.5 Consequently, the need to develop innovative approaches to peer review is increasingly clear. In this context, the emergence of artificial intelligence (AI) presents both opportunities and dilemmas for scholars, reviewers, editors, and societal stakeholders. For example, recent findings8 from a survey of more than 5000 researchers revealed that nearly half consider it acceptable to involve generative AI in peer review, and most support the use of AI tools to assist reviewers. Yet, 60% oppose the idea of AI drafting full review reports as a replacement for human expert-generated reports, citing concerns about confidentiality and trust. These findings highlight novel ethical tensions for journal editors, who rely on peer review to help inform the evaluation of scholarly work. As reviewers experiment with AI tools to ease workload and improve efficiency, the ways in which AI is used by peer reviewers have broad implications. At a systems level, for instance, the unfettered use of AI in peer review could introduce unvetted or potentially erroneous material into the same databases that AI models draw upon, compounding misinformation and potentially undermining the credibility of scientific communication for society at large.9 In addition, issues of author consent, protection of identifiable data, intellectual property rights, and other ethical safeguards remain undefined.10 These challenges thus raise important questions for all stakeholders in the development and publication of scholarship, including authors, reviewers, editors, readers, and the public. Might there be appropriate, narrowly defined, and practical applications of AI that enhance, rather than erode, the peer-review process? What practices can be put in place to ensure that engagement of AI facilitates well-informed, rigorous, and fair-minded reviews? Could AI help reduce reviewer fatigue, augment expertise, enhance feedback quality, and strengthen scholarly work? Or will its use compromise confidentiality, introduce bias, and ultimately erode the trust that underpins the peer-review process? Given the rapid evolution of AI and the novel challenges it presents for academic publishing, we at Academic Medicine and MedEdPORTAL have adopted a principle-driven approach to guide the use of AI in peer review. The principles of professional accountability, confidentiality, fair-mindedness and avoidance of harm, and transparency serve as the framework for our approach. The journals’ policy on AI use for peer review appears in Box 1, with examples of appropriate and inappropriate uses of AI provided in Table 1. Our aim in developing this framework is to anchor reviewer practices in shared values, enabling reviewers to thoughtfully adopt tools that are genuinely useful while preserving the integrity and role of expert peer review. Our intention is thus not to prohibit the use of AI, but rather to provide clear guidance and structure for peer reviewers and the range of stakeholders of our journals so that its integration enhances, rather than compromises, the rigor and trust that define our scholarly work. Examples of appropriate and inappropriate uses of artificial intelligence (AI) for peer reviewers. Examples of appropriate and inappropriate uses of artificial intelligence (AI) for peer reviewers. Peer reviewers are accountable, as individuals with specialized expertise, as members of a profession, and as human beings, for the quality and integrity of their evaluations of their colleagues’ scholarship. Concretely, this accountability means that reviewers should comment only on scholarship that falls within their area of expertise and refrain from commenting on work outside their scope of knowledge. For example, peer reviewers may find it useful to use AI tools to retrieve recent literature to inform or strengthen their understanding of a topic. However, AI should not be used as a substitute for a reviewer’s own scholarly judgment or disciplinary expertise. As members of a profession, peer reviewers are expected to conduct their reviews in a manner that respects the imperative for excellence and trustworthiness in science and scholarship more broadly. This professional expectation includes demonstrating respect for their colleagues whose work is being reviewed. In this context, some have advocated a “golden rule” approach in which reviewers should only use AI in ways that they would hope and expect others to use AI when evaluating their own work.10 And, while it may seem self-evident, it is important to emphasize that peer reviewers are humans. An AI tool cannot assume, or be assigned, professional accountability.11 Just as authors must be human, so too must peer reviewers. To demonstrate professional accountability, peer reviewers are expected to engage directly with manuscripts under review and to apply their own scholarly judgment in formulating comments and recommendations. They should limit AI assistance to acceptable uses (see Box 1) that do not involve uploading confidential materials or outsourcing evaluation. Reviewers must accept full responsibility for the accuracy of their reviews submitted and be prepared to address editor questions about the origin of their comments or any practices that could limit or bias the review. Importantly, professional accountability includes being truthful in disclosing whether and how AI was used in their conduct of the peer review. Academic Medicine and MedEdPORTAL adhere to the Committee on Publication Ethics12 and the International Committee of Medical Journal Editors13 guidance regarding peer review. This guidance is unequivocal. Reviewers must treat all author materials (eg, abstracts, figures, full texts) under review as strictly confidential and not share them without permission. Confidentiality is essential to the integrity of peer review because it protects authors’ unpublished ideas and data, preserves trust in the review process, and protects intellectual property from exposure or misuse. In simplest terms, just as a physician should not read a patient’s clinical records aloud in an elevator, a peer reviewer should not share the contents of an unpublished manuscript with others. The same standard applies in digital or physical environments, where confidentiality could be compromised. Authors submit their manuscripts to a journal with the expectation that the submission materials will be treated as confidential. This principle of confidentiality extends to the use of AI tools. Uploading any portion of a manuscript under review into an AI tool risks breaching confidentiality. Such actions risk exposing authors’ intellectual property to third-party servers where it may be stored, accessed by others, or used for future model training. There can be no guarantee of how AI tools will use materials once they have been uploaded, which carries the potential for misuse of author materials and the introduction of weak or biased data, analyses, and interpretations into public models that will serve as the basis of subsequent AI-generated content. While some institutions offer “closed” systems, which claim not to train on uploaded content or feed it into public models, they can also pose risks (eg, internal data circulation or system logging). For this reason, reviewers must not upload author materials under review to any AI tool, whether it is closed or open. Peer review should be based on expertise and fair-minded assessment and should not be delegated to AI tools. While AI may help with mechanical tasks like clarifying grammar, organizing notes, or summarizing a reviewer’s own comments, the intellectual work of assessing a manuscript’s originality, significance, methodology, and implications of the work are nontransferable responsibilities and rest with the human reviewer. Reviewers should approach their work in a neutral and balanced manner and seek to minimize sources of bias that may affect their assessment of scholarship. Fair-mindedness in peer review is an intentional effort to minimize bias. It is essential as a safeguard against harm that may occur when flawed or erroneous work is published or when contributions are undervalued due to implicit bias or uneven standards. Related to this, because AI tools are trained on uneven corpora, they may perpetuate or amplify existing biases in scholarship,14–17 threatening a fair-minded evaluation of a manuscript. In relying on AI, peer reviewers may introduce distortions, including bias from underlying datasets as well as erroneous information and AI “hallucinations.” To uphold fairness and avoid harm, peer reviewers should be cautious in how they engage with AI-generated content. They must be on guard for ways in which AI might shape their judgment, whether by skewing it due to counterfactual information or errors of emphasis or by disadvantaging particular research topics, methodologies, authors, or populations under study. Transparency is essential in maintaining trust and accountability in peer review. Peer reviewers have a responsibility to disclose whether and how AI tools contributed to a review. Clear disclosure allows editors to interpret a review in its full context and maintains trust in the fairness and rigor of the process. While transparency is always appreciated, not all uses of AI require disclosure, though this may evolve as AI tools and related publishing standards continue to develop. For example, currently, routine language editing (eg, spell check, grammar check) performed by AI tools does not require disclosure. More substantive uses of AI tools, however, such as rephrasing for clarity, generating an outline, summarizing publicly available information, or translating text, may be acceptable when appropriately disclosed to the editor. On the other hand, certain uses of AI, even when disclosed in a transparent manner, are never acceptable. For example, uploading confidential author materials in AI systems or delegating evaluative judgments to AI violates both confidentiality and accountability. Such unacceptable practices can compromise the trust that the scholarly community places in its reviewers and that society places in science. As editors and reviewers navigate the evolution of AI tools in academic publishing, the challenge is not merely to decide whether AI should be used in peer review, but how AI can be integrated in ways that uphold excellence in scholarship and trustworthiness in academia. The promise of AI is undeniable. When used appropriately, it aligns with the core purpose of peer review: to improve scholarship through thoughtful, expert evaluation. At the same time, the dilemmas posed by AI use in peer review are profound. Relying on AI to generate language or analysis introduces uncertainty about confidentiality, authorship, accountability, and bias. These concerns underscore the need for transparency and editorial oversight to ensure that peer review remains both rigorous and trustworthy. Together, the principles of professional accountability, confidentiality, fair-mindedness and avoidance of harm, and transparency provide a framework for responsible AI use in peer review. As AI tools continue to evolve, application of these principles will ensure that peer review remains rooted in human expertise, judgment, and integrity. Moving forward, we can imagine a future in which authors may opt to allow their manuscripts to be securely processed by journal-controlled AI systems, enabling both higher-quality and more sustainable peer review. Properly designed, such systems might help reviewers identify inconsistencies, highlight relevant literature, or check adherence to methodologic rigor and reporting guidelines—while remaining under the supervision of humans who are accountable for the work. We recognize our own responsibility to remain informed as the field evolves. Along the way, all journals must remain committed to the same principles that should guide reviewers. As AI tools are developed and tested to contribute to peer review in ethical, legitimate ways, editors always remain accountable for every editorial decision rendered. Author materials must continue to be protected as confidential, and authors must have the opportunity to opt into or out of any AI tools used by journals. To ensure fair-mindedness and avoidance of harm in editorial decision-making, editors will need to understand the potential weaknesses or biases of AI tools intended to support peer review and actively work to mitigate these. Finally, editors must be transparent with authors and reviewers about any changes to policies or practices about the use of AI in peer review. The availability and sophistication of these tools are quickly evolving, and journal editors must recognize when it is time to consider principle-driven changes to existing approaches. As AI tools become more sophisticated and embedded in scholarly work, journals will need to clarify expectations and provide equitable access to approved, secure tools to avoid creating new divides between reviewers who have access to advanced systems and those who do not. Realizing these possibilities will require not only technological innovation but also shared ethical commitments. For now, however, the responsible path is clear: AI tools may have narrow uses to help reviewers express their assessments of manuscripts but must never replace the expertise, judgment, and trust that define the peer-review process. Use of AI Tools in Peer Review: Academic Medicine and MedEdPORTAL Policy Reviewers have professional obligations regarding AI use, disclosure, and attestation. Author materials (eg, manuscripts, abstracts, figures, tables, appendices) under review are confidential and remain the intellectual property of the authors. Just as reviewers must not share any portion of these materials with any person, except in rare circumstances and only with the permission of the editor-in-chief, reviewers must not upload materials from authors into an AI system. Accordingly, the use of generative AI tools (eg, ChatGPT, Claude, Gemini) to analyze materials and/or to generate the contents of a review is prohibited. Authors should not use generative AI tools as a replacement for their own scope of expertise, knowledge, or judgment relevant to their evaluation of the materials. The journals recognize that some reviewers may wish to use AI tools to improve the grammar or readability of their own review. This use is permitted, provided the reviewer uploads only their own original work (ie, their own reviewer comments) and no identifiable information from the authors’ materials is included. Upon submitting a review, reviewers are required to disclose whether they have used AI tools in preparing their review. Reviewers are required to attest that no author materials were uploaded to AI tools, including publicly available as well as private or institution-based AI systems, and that the substance of their review was not generated by such tools. Only a human reviewer can be accountable for a submitted review, and reviewers are required to attest that they accept responsibility for the accuracy and integrity of their review. K. Fischer is an associate editor of Academic Medicine. M.B. DeVilbiss is on the editorial staff of Academic Medicine and MedEdPORTAL. L.A. Maggio is editor-in-chief of MedEdPORTAL. L.W. Roberts is editor-in-chief of Academic Medicine. The opinions expressed in this editorial do not necessarily reflect the opinions of the AAMC or its members.