Policy consensus, disagreements, and gaps regarding the use of artificial intelligence in authors, reviewers, publishers, editors, readers, and scholarship dissemination: a narrative review focus on medical scholarly publishing, analyzing attitudes of 6 institutions, 8 publishers, 139 journals and hundreds of articles

Abstract

Background and Objective

Artificial intelligence (AI) integration in medical publishing has improved efficiency but raised ethical controversies. Lacking systematic comparisons on AI-use polices across roles, this narrative review aims to summarize existing consensus, major disagreements, and policy gaps to guide future AI policy updates.

Methods

We comprehensively searched English and Chinese literature from 2016 to Jan 12, 2026 in Web of Science, PubMed, Scopus, EMBASE, and CNKI, and analyzed AI polices from 6 major organizations up to Jun 15, 2026, 8 top publishers up to Mar 1, 2026, and 139 stratified randomly sampled journals up to Apr 10, 2026, covering 25 publishing sub-scenarios.

Key Content and Findings

(I) Core consensus: clear agreement exists on confidentiality, disclosure of AI use, human verification and accountability, and the ineligibility of AI authorship. (II) Authors: AI is permitted for language polishing, translation, formatting, structural layout, literature searching, methodology, and study design. However, substantial disagreements remain regarding data collection, extraction, analysis and statistics, figure/table preparation, and manuscript writing among institutions, publishers, and scholars. Idea generation and downstream activities like revision and response letter drafting remain policy gaps. (III) Reviewers: AI is strictly prohibited from independent decision-making. However, disagreements persist regarding using AI to generate or optimize review comments. Checking revisions and response letters remains a policy gap. (IV) Editors/publishers: large policy gaps exist. Limited polices tend to prohibit AI from independent decisions while permit to select reviewers and draft/refine editorial comments. However, substantial disagreements remain over using AI for technical/compliance checks and content/substantive review. (V) Readers/dissemination: policies are largely absent. Available rules show a positive attitude toward readers using AI to translate and exact key points, and disseminators using AI to generate social media content and visualizations.

Conclusions

While current AI policies have reached consensus in many scenarios, substantial gaps and disagreements for editors, publishers, and readers must be prioritized in future updates. Future revisions should enhance specificity, scenario-based design, clarity, and accessibility.

Introduction

Background

Artificial intelligence (AI) technologies represented by machine learning and deep learning have developed rapidly in multiple fields and widely used in scholarly publishing. The launch of Chat Generative Pre-trained Transformer (ChatGPT) further marked the entry of generative artificial intelligence (GAI) into a phase of exponential growth (1). GAI is no longer limited to simple text translation. It can assist researchers in generating ideas and optimizing content structure, and substantially improve research efficiency in areas including literature reviews and data management (2). However, its application also brings many challenges, including increased risks of academic misconduct, such as AI plagiarism that is difficult to identify (3); blurred definitions of authorship (4); and frequent publication ethics issues (5). Against this background, stakeholders in scholarly publishing, including editors, reviewers, authors, and readers, have not yet reached consensus on AI-use scenarios, and their attitudes differ markedly (6). Some researchers have expressed concern about the use of AI in research, arguing that it may lead to cognitive dependence, thereby affecting the long-term quality of education, and may also generate large volumes of low-quality scholarly content (7,8). By contrast, other researchers believe that opportunities and challenges coexist and advocate actively accepting AI because it has immeasurable practical value in providing intelligent, multifunctional, and round-the-clock collaboration (9-11).

Rationale and knowledge gap

To address these challenges, major global publishers and journals, such as Springer Nature, Elsevier, and JAMA, have successively issued guidelines on AI use (12-16). Authoritative bodies, including the Committee on Publication Ethics (COPE) and the International Committee of Medical Journal Editors (ICMJE), have also issued relevant policies on AI use (17,18). However, these policies are numerous, and requirements differ across publishers, journals, and even institutions, creating major challenges and confusion for stakeholders in understanding and applying them. It is therefore necessary to conduct a review to clarify these policies. Although many previous reviews and studies have summarized or investigated the general institutional requirements regarding the boundaries of AI use and disclosure across journals or institutions, most have focused only on a single perspective, such as authors, reviewers, or editors (19-25), or on a single medical specialty, such as ophthalmology or cardiovascular medicine (26,27). In addition, no review appears to have systematically compared policy consensus and disagreements across institutions, publishers, and journals regarding AI use by various stakeholders and across detailed stages of the full medical scholarly publishing process.

Objective

This narrative review aimed to investigate the consensus, disagreements, and gaps in policy requirements for AI use across specific scenarios and stages of medical scholarly publishing involving authors, reviewers, editors, readers, and scholarly dissemination, as reflected in academic institutions, publishers and journals, and medical scholarly research contexts. On this basis, it sought to propose directions for future updates to AI policies. We present this article in accordance with the Narrative Review reporting checklist (available at https://atm.amegroups.com/article/view/10.21037/atm-2026-0169/rc).

Methods

This narrative review was structured around 25 sub-scenarios covering the use of AI throughout the scholarly publishing process, involving authors, reviewers, editors and publishers, readers, and scholarly dissemination. Table 1 lists these sub-scenarios and their corresponding Results subsections. The classification of each of the 25 sub-scenarios as consensus, disagreement, or a policy gap was performed independently by four authors (S.X., B.S., Y.H., and F.J.), after which the classifications were compared and consolidated. Any discrepancies among the four authors were resolved through discussion with K.Z. until consensus was reached. Positions were compared within source categories for each sub-scenario. “Consensus” was defined as an aligned position explicitly expressed across all relevant sources within a given source category; substantively divergent positions were classified as “disagreement”; and sub-scenarios that were not addressed in the reviewed documents or for which no explicit position was stated were classified as a “policy gap” or “policy silence”.

Table 1. Summary table of the 25 sub-scenarios of AI use.

No.	Perspective	Sub-scenario	Reported on section
1	Author	Authorship and disclosure	Sub-scenario 1. AI authorship and disclosure requirements
2	Author	Verification and accountability	Sub-scenario 2. Requirements for authors to verify and take accountability for AI output
3	Author	Researching	Sub-scenario 3. Using AI among authors for researching, including idea generation, methodology and study design such as search-strategy development, literature searching, and data collection and extraction
4	Author	Analyzing	Sub-scenario 4. Using AI among authors for analyzing, including data analysis and statistics, visualization and figure/table preparation based on study data, artwork generation, and image/video processing
5	Author	Writing	Sub-scenario 5. Using AI among authors for writing, including overall outlines, abstract writing, introduction writing, methods writing, results writing, discussion writing, conclusion writing, reference preparation, and graphical abstract generation
6	Author	Editing	Sub-scenario 6. Using AI among authors for editing, including language polishing, translation, structural layout, and formatting optimization
7	Author	Publishing	Sub-scenario 7. Using AI among authors for publishing, including pre-submission screening, recommending suitable journals, drafting cover letters, revising manuscripts, and writing response letters
8	Reviewer	Confidentiality and disclosure	Sub-scenario 8. Confidentiality and disclosure requirements
9	Reviewer	Verification and accountability	Sub-scenario 9. Verification and accountability requirements
10	Reviewer	Decision-making	Sub-scenario 10. Using AI for decision-making among reviewers
11	Reviewer	AI-generated, AI-assisted or AI-supplemented comments	Sub-scenario 11. Using AI among reviewers to generate review comments or to assist/supplement review comments
12	Reviewer	Improving readability	Sub-scenario 12. Using AI among reviewers to optimize the readability of review comments
13	Reviewer	Checking revisions and response letters	Sub-scenario 13. Using AI among reviewers to check manuscript revisions and response letters
14	Editor/Publisher	Confidentiality and disclosure	Sub-scenario 14. Confidentiality and disclosure requirements
15	Editor/Publisher	Verification and accountability	Sub-scenario 15. Verification and accountability
16	Editor/Publisher	Technical and compliance checks	Sub-scenario 16. Using AI among publishers and editors to check manuscript technical and compliance issues
17	Editor/Publisher	Content review and decision-making	Sub-scenario 17. Using AI among publishers and editors for content or substantive review and decision-making
18	Editor/Publisher	Selecting reviewers	Sub-scenario 18. Using AI among publishers and editors to select reviewers
19	Editor/Publisher	Summarizing content; drafting/improving comments	Sub-scenario 19. Using AI among publisher' and editors to summarize manuscript content and draft/improve comments
20	Editor/Publisher	Language polishing and copyediting	Sub-scenario 20. Using AI among publishers and editors for language polishing and copyediting
21	Editor/Publisher	Format optimization and typesetting	Sub-scenario 21. Using AI among publishers and editors for formatting optimization and typesetting
22	Reader/Dissemination	Confidentiality, disclosure, verification, and accountability	Sub-scenario 22. Confidentiality and disclosure, verification, and accountability
23	Reader/Dissemination	Generating social media posts, news, and visualizations	Sub-scenario 23. Using AI among publishers or media to generate downstream content for social media, press releases, visualizations, and similar outputs
24	Reader/Dissemination	Extracting key points or translation	Sub-scenario 24. Using AI among readers to extract key points and directly translate articles during reading
25	Reader/Dissemination	Generating posters or slides	Sub-scenario 25. Using AI among scholars to generate conference posters and PowerPoint presentations

The 25 sub-scenarios encompass AI-use scenarios across the full scholarly publishing process, involving authors, reviewers, editors and publishers, readers, and scholarly dissemination. These sub-scenarios form the analytical framework of this review. AI, artificial intelligence.

Literature search, inclusion, and information extraction

This narrative review was conducted based on a structured literature search and an explicitly defined study screening process. Bibliographic databases were searched to identify published literature and historical policy information reported therein. We searched Chinese- and English-language literature related to the current status of AI in medical scholarly publishing, published from January 1, 2016 to January 12, 2026, in five databases (Web of Science, PubMed, Scopus, EMBASE, and CNKI). Search terms included “artificial intelligence”, “machine learning”, “ChatGPT”, and “scholarly publishing”, combined using Boolean logic. The search summary is shown in Table 2. The detailed search strategies are provided in Appendix 1.

Table 2. The search strategy summary.

Items	Specification
Date of search	January 12th, 2026 for literature; January 1, 2026 to March 1, 2026 for publisher policies; January 1, 2026 to April 10, 2026 for journal policies; January 1, 2026 to March 12, 2026 and June 15, 2026 for searching policies by organizations, ethics committees, and standards-setting bodies
Databases and other sources searched	Databases: Web of Science, PubMed, Scopus, EMBASE, CNKI
	Publishers: Elsevier, Springer Nature, Wiley, MDPI, Taylor & Francis, Oxford University Press, SAGE, and Frontiers Media SA
	Journals: 139 journals (see Appendix 3)
	Organizations: COPE, ICMJE, the World Association of Medical Editors, the International Association of Scientific, Technical and Medical Publishers, the Institute of Scientific and Technical Information of China, and the Chinese Medical Association
Search terms used	“Artificial Intelligence”, machine learning, ChatGPT, “Academic Publishing”, “Author”, “Reviewer”, “Editor”, “Reader”
	Detailed search terms are presented in Appendix 1
Timeframe	2016.1.1–2026.1.12 for literature; up to March 1, 2026 for publishers; up to April 10 for journals; up to June 15 for organizations
Inclusion and exclusion criteria (only for literature)	Inclusion criteria: articles on the current status of AI in medical scholarly publishing, including, for reviewers, authors, and readers, the current use of AI tools, perceptions of the advantages and risks of using AI tools, differences in understanding publication policies, mainstream AI tools, and the strengths and limitations of tool performance; and for journal editors, AI-generated content detection, scope checking, manuscript screening, decision support, formatting optimization, typesetting and proofreading, image optimization, public dissemination, and related topics; Chinese and English
	Exclusion criterion: full text unavailable
	Study types: Original Article, Review, Editorial, Letter
Selection process (only for literature)	B.S. and S.X. screened titles and abstracts; Y.C. and Y.H. independently read the full texts to decide whether to include each article; when disagreements arose during the process, K.Z. joined the discussion until consensus was reached

AI, artificial intelligence; ChatGPT, Chat Generative Pre-trained Transformer; COPE, Committee on Publication Ethics.

Eligible article types were limited to Original Article, Editorial, Letter, and Review. Eligible articles focused on the current status of AI in medical scholarly publishing. First, for reviewers, authors, and readers, we summarized information on current AI use, perceived advantages and risks, and differences in understanding of publication policies. Second, for journal editors, we summarized AI-generated content detection, scope checking, manuscript assessment, decision support, formatting optimization, typesetting and proofreading, image optimization, public dissemination, reviewer identification, and related topics. B.S. and S.X. screened titles and abstracts. Y.C. and Y.H. independently read the full texts to decide whether to include each article. When disagreements arose during inclusion, K.Z. joined the discussion until consensus was reached. Articles for which full texts could not be downloaded were excluded, and the number and reasons for exclusion were recorded. The detailed literature inclusion process is shown in Figure 1. EndNote 21 was used for reference management. B.S., S.X., Y.H., and F.J. extracted standardized information from the included literature, including article type, target user (author/reviewer/editor/reader and dissemination), application scenarios, policies/guidelines mentioned, and core views/attitudes. The information extraction table is provided in Appendix 2.

Figure 1.

Flow diagram for narrative review.

Policies of publishers and affiliated journals on AI

The top 8 medical or general publishers by 2025 publication volume (search date: May 15, 2026), namely Elsevier, Springer Nature, Wiley, MDPI, Taylor & Francis, Oxford University Press (OUP), SAGE, and Frontiers Media SA (Frontiers), were included as information sources. Using IBM SPSS Statistics 31.0, we conducted stratified random sampling of 2,717 medical journals from these 8 publishers according to JCR quartiles. Five journals were further randomly selected from each JCR quartile for each publisher, if a publisher had fewer than 5 journals in each JCR quartile, all journals in that quartile were included. A total of 139 medical or general journals from these 8 publishers were ultimately included.

K.Z., F.J., and Y.C. manually searched the official websites of 8 publishers (28-35) and 139 journals (see Appendix 3), independently extracted policies related to AI use by authors/reviewers/editors/readers in different stages, and compared whether journal policies were consistent with those of their parent publishers. Where discrepancies emerged during this comparison, they were resolved through discussion and consensus among all authors. Publisher policies were searched and extracted from January 1, 2026 to March 1, 2026, and journal policies were searched and extracted from January 1, 2026 to April 10, 2026. The detailed information extraction table is provided in Appendix 2.

AI policies of scholarly publishing organizations

ZKP searched and extracted AI-use policies from organizations, ethics committees, and standards-setting bodies in scholarly publishing, including COPE, ICMJE, the World Association of Medical Editors (WAME), the International Association of Scientific, Technical and Medical Publishers (STM), the Institute of Scientific and Technical Information of China (ISTIC), and the Chinese Medical Association (CMA) (16-18,36-43). Policies were searched and extracted from January 1, 2026 to March 12, 2026, with an updated search conducted on June 15, 2026. The detailed information extraction table is provided in Appendix 2.

Policies and requirements on use of AI among authors

Sub-scenario 1. AI authorship and disclosure requirements

Regarding AI authorship, the latest regulations from the 6 institutions, 8 publishers, and 139 journals reviewed showed a consistent attitude: all clearly stated that AI does not qualify for authorship and cannot be listed as an author. At present, most scholars also oppose AI authorship. Opposition to AI authorship is mainly based on the fact that AI does not meet authorship criteria, including its inability to take accountability for the work or resolve conflicts of interest, and its lack of cognitive intentionality and inability to provide critical analysis. Aorigele Bao et al. noted that treating AI as a co-author fundamentally undermines the essence of human authorship and ignores human dignity and humanization (44). For example, Nurse Education in Practice published an article in 2023 that listed ChatGPT as a co-author, which triggered considerable controversy. The subsequent correction removed ChatGPT from the author list and retained only human authors (45). Juan Gorraiz’s bibliometric study also found that cases in which AI was listed as an author mostly occurred before 2023, whereas the current trend is to cite or disclose the AI tools (4).

However, regarding disclosure of AI use, the 6 institutions, 8 publishers, 139 journals, and the literature differed in their attitudes toward the timing, location, and content of disclosure. In terms of timing, among the 6 institutions, only ICMJE and STM emphasized that disclosure should be made in the cover letter at the initial submission stage. Among the 8 publishers, only OUP and MDPI required authors to declare AI use to editors at submission, whereas the remaining 6 publishers did not require a similar disclosure time point. Views in the literature tended to favor disclosure at submission. For example, Fuster et al. described in detail what should be disclosed in the cover letter, including the AI model, version, manufacturer, and the specific nature of its contribution, such as study design, implementation, analysis, writing, and reporting (14,46,47). However, some scholars expressed concern about overly early disclosure. Ozkara et al. noted that editors were clearly biased in favor of articles they believed were written by humans during assessment, and that advance disclosure of AI use might affect editorial decisions. Therefore, they suggested that authors disclose AI-tool use after the manuscript decision (48). In terms of disclosure location, the 6 institutions and 8 publishers generally specified that AI disclosure should appear in the Methods or Acknowledgments, with only Elsevier requiring a separate AI declaration section in the manuscript. Views in the literature were also largely consistent, with the Methods section focusing on dimensions of AI use related to study design and conduct, and the Acknowledgments focusing on dimensions related to manuscript writing and editing assistance (14,46,47). Buchman et al. further emphasized that concealing AI use that should have been disclosed in the Methods section violates scientific integrity and should be regarded as academic misconduct (49). Regarding disclosure content, ICMJE requires disclosure of the AI name, version, and prompts, while WAME, ISTIC, and CMA further emphasize disclosure of the specific time when AI was used. In addition, ISTIC and CMA emphasize disclosure of the specific AI-use process or the specific stage in which AI participated, and they state that AI-generated ideas and viewpoints should be clearly marked to distinguish them from human ideas and viewpoints. CMA also explicitly requires disclosure of whether humans have reviewed AI-generated content and disclosure that humans bear accountability for the truthfulness and accuracy of all content. Publisher requirements show similarly marked differences: Frontiers requires disclosure of the AI tool’s name, version, model, and source; SAGE and MDPI require disclosure of the stage and details of use; OUP and Taylor & Francis require disclosure of the method and reason for use in addition to the name and version; and Springer Nature only mentions disclosure without defining the required content. In addition, SAGE, MDPI, Wiley, and Elsevier define circumstances in which disclosure is not required, whereas the remaining 4 publishers do not have corresponding provisions. OUP and Elsevier further emphasize disclosure of authors’ supervision and verification of AI output. Wiley’s disclosure requirements also mention authors’ roles in guiding or reviewing AI-related work, how privacy and compliance were ensured, and the use of supplementary materials to ensure transparency and reproducibility of AI use. Most journal disclosure requirements were consistent with those of their parent publishers, but some journals differed. For example, disclosure requirements in the 19 sampled OUP journals were all inconsistent with OUP requirements. Specifically, these journals required disclosure only of the method of use and the tool, rather than requiring disclosure of the reason for use and whether authors performed verification steps as specified by OUP. Another 6 OUP journals had no AI policy, no disclosure requirement, and no link to the OUP policy. Similarly, 2 Elsevier journals had no AI policy or disclosure requirement and no link to the Elsevier policy, while 3 other journals had requirements that differed from Elsevier’s. The literature has also provided a logical summary of disclosure details. For example, Buchman et al. stated that disclosure mainly includes the purpose of use (Why), the augmented intelligence used (What), the method of use (How), and which author evaluated the results (Who) (49).

In summary, the academic community has reached consensus on AI authorship, and there is also basic consensus on the location of AI disclosure. However, substantial disagreements remain regarding the timing and content of AI disclosure. Journal requirements and those of their parent publishers differ in this regard.

Sub-scenario 2. Requirements for authors to verify and take accountability for AI output

Regarding the need for authors to verify and take accountability for AI output, the 6 institutions, 8 publishers, 139 journals, and the literature reviewed showed a consistent attitude: authors are ultimately accountable and must check and verify the truthfulness, accuracy, relevance, validity, and originality of AI output or recommendations. For example, ICMJE specifically states that authors should carefully review text and images, including whether plagiarism exists, whether cited materials are appropriately cited, whether references are fabricated, and whether content is biased, because AI may generate content that sounds authoritative but is wrong, incomplete, or biased. Hosseini et al. provided the rationale behind this requirement, including that AI responses are generated according to researchers’ prompts and requests rather than initiating text generation independently; that authors are the agents who combine or revise AI-generated text with their own writing; and that authors are the people who ultimately receive credit for the text (50).

Sub-scenario 3. Using AI among authors for researching, including idea generation, methodology and study design such as search-strategy development, literature searching, and data collection and extraction

At the starting point of authors’ use of AI for researching, namely idea generation, most institutions, publishers, and journals provide no information, making this a policy gap. This includes 4 of the 6 institutions and 7 of the 8 publishers. The limited available information is also inconsistent and contradictory. Specifically, although STM does not mention idea generation separately, it states that authors may use GAI only as a basic tool and for text/document polishing, correction, formatting, and editing. Any use beyond these scenarios must be fully disclosed and judged according to journal policy and editorial judgment. It can therefore be inferred that STM neither explicitly prohibits nor explicitly permits such use, but emphasizes disclosure, journal rules, and editorial judgment. CMA also does not explicitly mention idea generation, but it states that AI use is prohibited when important intellectual labor is involved, suggesting a reserved attitude. However, a contradiction also exists here, because CMA’s disclosure requirements state that AI-related disclosure should also mark AI viewpoints, which implies that such use is not prohibited, because marking would not be needed if it were prohibited. Taylor & Francis states that authors may use AI for creative ideation and exploration in settings that ensure high standards of data security, confidentiality, and copyright protection, and that authors must take full accountability for this use. From journals’ perspective, 19 Taylor & Francis journals were consistent with their publisher’s requirements. However, 3 Elsevier journals explicitly stated that AI may be used only to improve readability and language, although they did not separately mention other permissible uses, it can be inferred that these 3 journals do not support using AI for idea generation. Similarly, the current literature shows major disagreement among authors regarding AI-generated article ideas. This is reflected in the fact that, while AI use for generating research ideas is recognized as helpful for identifying scientific objectives and research gaps, and while AI can help researchers generate research questions, identify overlooked topics, connect different areas of literature, and generate innovative research ideas that differ from conventional thinking by simulating human thought and creativity and analyzing large volumes of literature and data, there are also concerns that long-term use may lead to algorithmic uniformity and produce increasingly similar research that caters to hot topics (51-53). Some scholars believe that AI should be used as an amplifier of innovation rather than the sole source of innovation. Specifically, individuals may first brainstorm ideas independently, then use AI for expansion, proceed to group discussion, and finally conduct another round of individual brainstorming to form an article idea (53).

In terms of methodology and study design, such as search-strategy development, both ISTIC and CMA permit such use. STM again neither explicitly prohibits nor explicitly permits it, but emphasizes disclosure, journal rules, and editorial judgment. ICMJE, COPE, and WAME provide no relevant information. MDPI states that AI may be used for study design. Wiley provides more detailed definitions than MDPI: for example, when AI is used to generate questionnaires based on research objectives or to develop or extend experimental protocols by analyzing existing methods, it may be used but must be disclosed, whereas when AI is used to organize protocols or explain research-method concepts, it may be used without disclosure. Although SAGE does not explicitly state whether AI may be used for study design, it explicitly prohibits substituting generative AI tools for participants in qualitative research interviews and prohibits using AI for experiential and thematic analysis. The other 5 publishers and the vast majority of journals provide no relevant information. The literature generally supports such use, including using AI to help design scale items and plan statistical analyses, but emphasizes that authors should be alert to “optimization hallucination”: AI may design a perfect protocol, but that protocol may not be suitable for real-world application. For example, AI can serve as a virtual interactive object to simulate question-and-answer data generation, but by design this can only be used as an exploratory and method-optimization tool in the early trial stage and cannot replace real human experience (53). YI et al., after studying policies from many medical journals, also concluded that authors may use AI during the study-design stage to help summarize literature, optimize study protocols, and assist clinical research, while emphasizing that all of these steps require human verification and evaluation (54).

Regarding searching in actual researching activities, the policy positions of the 6 institutions were similar to those described above for methodology and study design. Among the 8 publishers, SAGE supports AI assistance in compiling relevant materials, and Taylor & Francis supports interactive online retrieval using search engines enhanced by large language models. Wiley supports AI independently generating or executing search strategies, including designing complete search strings, running database queries, or prescreening results for literature reviews or systematic reviews. Wiley also supports applying inclusion/exclusion criteria to screen or select studies for systematic reviews or meta-analyses (all requiring disclosure), and using keywords or phrases to assist searches and recommend relevant resources (requiring authors to review and decide whether to include the final results, without requiring disclosure). Except for a very small number of journals, the vast majority of journal policies were consistent with those of their parent publishers. Literature searching and screening are typically repetitive and time-consuming, and published literature has affirmed that AI use at this stage can substantially save authors’ time. For example, Ciule et al. compared Scopus AI, a search tool developed by Elsevier, with traditional literature retrieval tools and found that, for researchers with limited knowledge and understanding of a topic, Scopus AI can serve as a valuable tool for obtaining broad perspectives and discoveries. Traditional keyword searching provides precise and manageable results, which is advantageous for researchers who are already familiar with the topic and are seeking specific studies. Combining the two methods can offer a balanced approach, with Scopus AI used for an initial broad search followed by traditional keyword searching to obtain more refined results (55). One study evaluated AI for literature screening and extraction in systematic reviews and found that, compared with human screening alone, screening combined with AI saved an average of 45.4% of workload, while AI screening achieved a sensitivity of 95.8%. The authors ultimately supported using AI as a second reviewer for abstract and title screening during updated searches (56). However, published studies have also expressed concern about AI-assisted literature searching, mainly because AI cannot detect whether a journal is predatory, that is, a journal that charges authors high fees while providing little peer review, and therefore cannot evaluate the quality of the retrieved literature (57). Scholars therefore state that, when using AI for literature searching, authors must verify and evaluate the quality of the retrieved literature.

When data collection and data extraction are involved, the policies of several institutions begin to change. The main change is that both STM and CMA take a more cautious position and explicitly prohibit the use of AI for data collection and extraction. SAGE takes a similarly cautious attitude and explicitly prohibits using AI to generate data through a series of prompts or to submit data. OUP requires that, if authors use GAI for data collection, they must submit it as part of the materials and obtain OUP’s prior written permission. By contrast, MDPI explicitly permits AI use for data generation or data collection. Wiley also permits, with transparent disclosure, the use of AI to create databases, patient registries, or research recruitment platforms, or to extract references or metadata from papers or databases for literature reviews. The remaining 4 publishers, as well as many journals including sampled OUP journals, have not issued relevant requirements. Attitudes in the literature are similarly polarized. On the one hand, scholars recognize that AI-assisted data extraction can improve efficiency and construct knowledge graphs to help researchers understand field hotspots clearly and intuitively (58). On the other hand, when using AI to collect data, scholars also emphasize restricted use: the truthfulness and reliability of extracted data must be manually checked, including in scenarios such as literature research, medical concept explanation, opinion-based material research, and clinical data collection. When clinical data are involved, medical ethics must also be followed, such as strictly prohibiting the illegal acquisition or use of patient data (54,59). Furthermore, when GAI is used to collect data from human participants, informed consent from participants must be obtained (60). Helms Andersen et al. also showed that, compared with human extraction, AI tools performed well and similarly in systematic-review data extraction, especially for standardized variables. However, error analysis showed that 4% of data points were fabricated by human intervention. They therefore proposed using AI-assisted extraction to replace the second human data extraction, with a second human data reviewer comparing and reconciling data extracted by one human and one AI tool (26).

Taken together, regarding authors’ use of AI for researching, there are extensive policy gaps in the area of idea generation, and the limited existing requirements show many inconsistencies and contradictions. Substantial disagreements exist in institutions and publishers, although attitudes in the literature tend to be relatively accepting of a collaborative-assistance model. Requirements for data collection and extraction show similarly large disagreements. For methodology and study design, such as search strategies, except where specific types of design are prohibited, AI use is generally supported under the premise of human verification and vigilance against optimization hallucination. The consensus status of policy requirements for literature searching is broadly similar.

Sub-scenario 4. Using AI among authors for analyzing, including data analysis and statistics, visualization and figure/table preparation based on study data, artwork generation, and image/video processing

Regarding data analysis and statistics, except for STM, which explicitly prohibits AI participation, the other 5 institutions either permit it or do not explicitly prohibit it. From the publisher’s perspective, corresponding information is either absent or AI use is permitted with or without restrictions. Specifically, OUP requires prior written permission from OUP and submission as part of the materials if GAI is used for data analysis. MDPI permits AI for data analysis or interpretation without further restrictions. Wiley provides more detailed requirements in this regard, permitting AI to process medical imaging or microscopy data for analysis; to code interview or textual data and identify patterns or themes in large datasets; to perform statistical analysis and flag outliers; and to interpret results, highlight relationships or trends in data, or synthesize findings from multiple datasets. Several other publishers provide no relevant information. From the journal’s perspective, except for more than 10 journals with no relevant policies, the remaining journal requirements were consistent with those of their parent publishers. It should be emphasized that AI is an analytic tool, not a substitute for methodological competence. Authors must still possess basic abilities in statistics, study design, and theoretical interpretation, and must conduct cross-validation through different methods, tools, and theoretical logic. GAI has the ability to analyze literature and data at scale and can help identify trends, patterns, and even dynamic interactions that humans cannot calculate, opening new paths for high-throughput scientific research that were previously difficult to imagine (61). However, in recommendations on statistical methods, data analysis, description of statistical results, and intelligent parameter annotation, AI should serve only as an auxiliary statistical-processing tool and should not replace researchers’ analysis. All analyzed data must be genuine and must never be tampered with or fabricated, and professional verification is required. Data interpretation, as the most central component, belongs to researchers’ professional judgment and AI use should be prohibited there to avoid misleading conclusions through erroneous interpretation (54). An initiative on standardized AI use in medical journals also proposed that AI may be used for statistical analysis based on real scientific data, including suggestions and guidance on the choice of statistical methods, but authors must be responsible for modeling methods and testing logic and must ensure the reliability of analytic results through independent verification (62).

Attitudes are more cautious regarding the use of AI to create visualizations and figures/tables based on study data, generate artwork, or process images and videos. In addition to STM’s explicit prohibition, CMA states that AI may not create, alter, or manipulate original data, including figures, audio, and video, and that AI may be used only to recommend figure types and assist figure preparation. Among major publishers, a few have a simply permissive attitude, while most allow such use to some extent under strict restrictions. Specifically, Frontiers and MDPI permit AI to generate or edit images without many restrictions. However, SAGE, Wiley, Springer Nature, and Elsevier all have explicit prohibitions on AI use in images and videos. Although SAGE states that AI may be used to generate representative illustrations or infographics, it explicitly prohibits presenting AI-generated images as unique or novel research images. Wiley states that AI may use underlying data to generate charts, graphs, curves, heat maps, scatterplots, or statistical figures; may generate or edit workflow diagrams, conceptual frameworks, experimental apparatus visualizations, or flow diagrams; and may create synthetic images, but these must be clearly labeled in the figure or legend, and original images must be provided at submission. Wiley explicitly prohibits using AI to generate, modify, or enhance factual and evidentiary images. Springer Nature permits AI-produced videos, animations, edited illustrations, cartoons, and similar materials when labeled “generated by AI”, but imposes strict restrictions on image sources. Elsevier permits AI only for adjusting image brightness, contrast, or color balance without obscuring or eliminating any information from the original image, and explicitly prohibits using AI to create or modify images, including enhancing, blurring, moving, removing, or adding specific features in images or figures. Authors must also describe in the Methods the specific way in which AI was used in image creation or modification and must provide the original images. In addition, although OUP permits AI use for figure and table production, some journals under OUP specifically removed this permissive policy. Other journal policies were generally consistent with those of their parent publishers. Views in the literature currently generally support AI use for creating data-based statistical visualizations, especially in systematic reviews and meta-analyses, because these are automated and repetitive tasks that present data in an easily understandable format and add value to research. For example, SciNote (63) can be used to organize data, enabling more accurate presentation in research papers and reducing the possibility of reporting errors (64). It is sufficient to provide detailed information about the generated content, such as the model or tool name, version number, extension, and manufacturer, and authors remain accountable for this generated content (46,65). However, the literature also states that AI-assisted generation of research-result images, such as experimental images or medical examination images, must be strictly prohibited. These images must be genuinely obtained, and AI generation constitutes data fabrication and must be absolutely prohibited. An exception exists when AI-generated content is part of the research method, for example, AI-generated images may be used as model-output examples in papers about image-generation algorithms (66). Some scholars even state that authors should not use images generated by AI tools or similar technologies in manuscripts unless they have discussed this with the editor-in-chief and obtained explicit permission (47).

In brief, regarding AI use for data analysis and statistics, many institutions and publishers lack established formal policies. Where relevant information exists, institutions tend to prohibit such use, whereas publishers tend to permit it, either with or without restrictions. Attitudes in the literature are generally consistent with those of publishers. For AI applications in visualization, institutional attitudes and publisher attitudes also differ substantially, reflected in more explicit prohibitions by institutions and permissive publisher positions under detailed conditions. In this area, journals also show relatively large inconsistencies with their parent publishers. Attitudes in the literature show slight fragmentation despite broad consensus.

Sub-scenario 5. Using AI among authors for writing, including overall outlines, abstract writing, introduction writing, methods writing, results writing, discussion writing, conclusion writing, reference preparation, and graphical abstract generation

Regarding rules on using AI for writing, including broad policies and specific rules targeting writing and partial manuscript drafting such as abstracts, introductions, methods, results, discussions, conclusions, references, the 6 institutions differ greatly. ICMJE, COPE, and WAME do not explicitly permit or explicitly prohibit such use. STM explicitly prohibits AI participation in writing the results section, while not explicitly prohibiting other sections and emphasizing disclosure, journal rules, and editorial judgment. Both ISTIC and CMA emphasize that AI may not be used to write an entire paper, and that important parts involving scientific and intellectual contributions should be completed by authors themselves, including the results, discussion, and conclusion. It is worth noting separately that STM and ISTIC both prohibit using AI to generate graphical abstracts. Among publishers, Frontiers, MDPI, and Wiley all permit AI-generated text under transparent disclosure and author verification; Frontiers also separately lists extracting references from papers or databases as a permitted AI use. SAGE specifically permits AI-assisted literature review and reference organization only when erroneous text, fabricated citations, and false statements are not generated, and does not mention broader writing rules. Elsevier separately mentions permission for AI-made graphical abstracts if authors obtain prior permission and ensure that the materials are compliant and the content is correct. Several other publishers do not explicitly state their positions, and journal information is generally consistent with that of parent publishers, except that some OUP journals permit AI-generated content even though OUP itself has no explicit permissive policy.

Studies in the literature provide more detailed attitude information. For abstracts, multiple studies have found no significant difference between AI-written and human-written abstracts, and many studies express support for AI use in generating abstracts, although they still recommend that AI assist in generating abstracts followed by human proofreading (67-69). However, this supportive attitude is not present for writing introductions, discussions, and similar sections, mainly because of problems with AI citing low-quality literature, citing fabricated literature, and generating inaccurate references. For example, Sheridan et al. used an AI application with standardized prompts to generate discussion and conclusion sections based on provided text, and found that although the time required for AI to write the discussion was far shorter than for humans, the median score for AI-written discussion and conclusion sections was only 60, lower than that for other human-written sections (70). Similarly, although one study found that GPT-4.0 produced introductions comparable to those written by humans in readability and content quality (71), subsequent research found that all references cited in introductions written by GPT-3.5 were fabricated (72). Giavina Bianchi et al. reported three scenarios in which academic papers were written by humans, AI, or both. Compared with human writing, combined writing and AI generation reduced writing time by 83.8% and 84.3%, respectively, and improved quality by 13.9% and 11.1%, respectively. AI demonstrated the ability to enhance multiple manuscript sections, including the title, short title, abstract, introduction, discussion, and conclusion, but AI manuscripts received lower reference scores (73). Aiumtrakul et al. compared the accuracy of references generated by ChatGPT, Bing Chat, and Bard AI and found that even ChatGPT, which had the highest accuracy, achieved only 38%. Among references generated by these AI tools, DOI was the most common inaccurate component, followed by author names, publication year/issue, journal title, and reference title (74). Even though GPT-4.0 had a much lower reference citation error rate than GPT-3.5, its error rate still reached 20.6% (75). In addition, some studies point out that GAI cannot write reliable medical articles because it lacks specific qualifications or deep understanding of medical issues. It can provide general information but cannot guarantee the accuracy and correctness of that information (76). Therefore, some studies do not recommend using generative AI for any prompt-based content generation (77). A further concern is that overreliance on AI writing will weaken basic writing and critical-thinking abilities.

Therefore, institutions with relevant policy information have largely reached consensus that AI should not be used to write an entire article or important sections involving intellectual and scientific contributions. However, many publishers grant broader permission for AI writing. The literature takes completely different recommendation positions for abstracts and non-abstract sections, mainly because of concerns about fabricated, low-quality, and inaccurate literature in AI writing, as well as concerns about the generality and correctness of AI-generated content.

Sub-scenario 6. Using AI among authors for editing, including language polishing, translation, structural layout, and formatting optimization

Regarding authors’ use of AI for editing, including language polishing, translation, and formatting optimization, institutions, publishers, journals, and studies show a highly unified consensus permitting such use. This includes editing grammar, structure, spelling, punctuation, formatting, and other aspects that do not involve substantive content. This has also received broad recognition in academia, because authors, editors, reviewers, and readers prefer to focus on content itself rather than English proficiency. In addition, AI can automatically adjust manuscript formatting and citation styles according to specific scholarly standards or journal requirements, further reducing workload and improving the efficiency of scholarly publishing (58). For example, when EndNote or Zotero is integrated with AI, it can suggest and format references according to specific journal requirements, saving substantial effort (64). AI tools such as ChatGPT and Grammarly have also been reported to play an important role in helping non-native English speakers improve manuscript grammar, style, and fluency (78). In terms of translation, scholars have also stated that AI translation can improve the readability of manuscripts from non-native English-speaking scholars (51) and reduce the burden on scholars in non-English-speaking countries when reading or reviewing research written in languages in which they are not proficient (61).

In addition, multiple institutions and publishers state that such AI use in many cases requires no disclosure whatsoever. For example, STM states that AI participation in language polishing and structural optimization does not need to be disclosed. CMA also clearly states that using AI to optimize and adjust manuscript formatting does not need to be disclosed. SAGE, Springer Nature, and Wiley state that language polishing and formatting optimization do not require disclosure. However, Wiley emphasizes that disclosure is required if AI is used for text translation, restructuring arguments, or reorganizing the writing structure. Some OUP journals also stipulate that AI-assisted translation requires disclosure. Some Elsevier journals require disclosure even when AI is used only to improve language readability. In addition, if AI editing involves optimizing the tone of the text, Springer Nature considers disclosure unnecessary, whereas Wiley still states that disclosure is required. WAME states that AI participation must be disclosed for all of the above stages.

It is worth noting that, regardless of whether disclosure is required, the literature further points out that AI use in manuscript editing still requires verification, because improper use may seriously contaminate scientific thinking (79). Stojanov also issued a cautionary note: although AI lowers geographic and language barriers and increases linguistic equity, it may also lead to homogenized writing in the future and reduce original text. Therefore, Stojanov suggests that we should not blindly accept rewrites or translations provided by large language models, but should instead use them as inspiration for self-corrective writing (80).

Taken together, consensus-based permission has been reached for using AI in editing, but requirements differ regarding when disclosure is needed. While enjoying the unprecedented benefits AI brings to language and formatting optimization, we must also remain alert to the risks of content homogenization and failure to verify output.

Sub-scenario 7. Using AI among authors for publishing, including pre-submission screening, recommending suitable journals, drafting cover letters, revising manuscripts, and writing response letters

Compared with the preceding stages, after a project and manuscript have been completed and the work enters the publishing process, the 6 institutions have many gaps in their rules on authors’ use of AI. Among the policy requirements of the 6 institutions, 8 publishers, and 139 journals, no information was found on whether authors may use AI for pre-submission screening, recommending suitable journals, drafting cover letters, or revising manuscripts after receiving external review comments.

Regarding whether AI may be used to draft response letters, only WAME states that disclosure is required if AI is used, and ISTIC states that, if a journal does not explicitly prohibit it, AI may assist in conceiving the framework of a response letter and organizing its content. ChatGPT and similar AI tools have also been found to have potential to assist authors in completing the peer-review process. They can help revise response letters step by step, including understanding reviewer comments, drafting preliminary responses, improving manuscript revisions, and ensuring compliance with journal guidelines (81). However, the literature emphasizes that authors must be responsible when responses to reviewers involve substantive rebuttals (82).

The literature supports the use of AI to recommend suitable journals before submission. Identifying relevant journals or conferences is an important step in publishing research, because excellent articles may be rejected by unsuitable journals or reach smaller readerships. By analyzing an article’s topic, keywords, and quality, as well as author needs such as review cycle and acceptance rate, GAI can recommend journals that precisely match the author’s research field and thereby improve the likelihood of success (83,84). Some articles also state that AI can serve as a simple pre-submission review tool to help authors assess whether their manuscript fits a journal’s scope and methodological requirements, thereby allowing authors to improve the manuscript before formal peer review (85).

In short, substantial policy gaps exist across institutions, publishers and journals regarding the use of AI during the publishing stages after the research project and manuscript are complete, while the literature generally reflects a favorable attitude toward its active adoption.

Policies and requirements on use of AI among reviewers

Sub-scenario 8. Confidentiality and disclosure requirements

Regarding confidentiality, ICMJE, STM, ISTIC, and CMA all explicitly prohibit reviewers from uploading manuscripts, supplementary materials, core ideas, or key data to any location that cannot ensure manuscript confidentiality, including publicly available AI platforms. Publishers and journals have a similar consensus-based strict prohibition. SAGE, OUP, and Elsevier explicitly and strictly prohibit such use, and OUP additionally prohibits reviewers from uploading all or part of project proposals and manuscripts to generative AI tools. SAGE and OUP also require reviewers to report to the publisher if they discover that authors have violated OUP’s AI-use policy. Frontiers, Taylor & Francis, MDPI, Wiley, and Springer Nature state that AI may be used under certain conditions, but all explicitly prohibit reviewers from uploading manuscript content to generative AI tools, except for internal AI tools provided by Frontiers. Leakage of manuscript information is one of the most concerning risks in AI-assisted review. When part or all of a manuscript is placed into an AI tool, the content may be cached by the large model and it may be unclear how long it will be cached, thereby infringing authors’ manuscript privacy and intellectual property rights. Therefore, prohibiting reviewers from uploading manuscripts to AI systems is a broad consensus among Chinese and Western journals (85,86). Multiple journals and institutions have issued explicit confidentiality policies: science prohibits the use of AI tools in peer review to prevent potential bias and misconduct (87). JAVMA or AJVR further emphasizes (77) that reviewers who agree to review a manuscript are deemed to have agreed to keep the text confidential and may not upload any part of the manuscript to generative AI tools. This is especially important in medicine, where manuscripts may involve patients’ private health information and therefore have higher requirements for data privacy compliance (88). In addition, de Bessa et al. suggest that reviewers should prioritize secure and reliable AI tools approved by their organizations and encourage the use of offline or locally hosted large language model (LLM) models (89,90).

Regarding disclosure, because STM and CMA prohibit reviewers from using AI for peer review, disclosure is not applicable. Except that no disclosure requirements for reviewers’ use of AI were found in COPE and ISTIC, the remaining ICMJE and WAME both explicitly state that reviewers must disclose AI use, including the method of use, to the journal, and WAME further mentions that authors should also be informed when reviewers use AI. Similarly, Frontiers requires reviewers who use external AI tools to obtain editorial approval and disclose to authors which AI tools were used. MDPI and Springer Nature require reviewers to disclose AI use when submitting peer-review reports. In addition, one study found that among the top 50 dental journals by impact factor, 92% of journals permitting AI use required mandatory transparent disclosure (90). Journal of Medical Internet Research also requires reviewers to describe in detail the specific application of AI in generating review reports and to thoroughly review the terms of use of GCAI tools to ensure confidentiality (91). JAMA policy also requires that, beyond basic checks of grammar, spelling, and references, reviewers who use AI tools must provide the tool’s name, version number, extension, and manufacturer, and indicate how it was used (14).

Sub-scenario 9. Verification and accountability requirements

ICMJE, COPE, and ISTIC all mention that AI-generated comments must be manually verified and reviewed for completeness, accuracy, truthfulness, and fairness. ICMJE and ISTIC also state that reviewers bear full accountability for the review comments they produce and must ensure the objectivity and professionalism of the review. Publishers have similar consensus-based requirements. Elsevier and MDPI explicitly state that reviewers are fully accountable for the content of the review report. Taylor & Francis requires reviewers to be responsible for the accuracy and completeness of review reports. Frontiers require reviewers to ensure the accuracy, completeness, and relevance of review reports and to take accountability for review comments, ensure the objectivity and professionalism of the review, and emphasizes that AI tools cannot replace reviewers in making review judgments. Among the 139 journals, except for some that do not mention this issue, all other journals have requirements consistent with those of their parent publishers. In the literature, scholars also state that accountability for reviewing manuscript content remains entirely with human reviewers (90). Reviewers should strictly comply with ethical guidelines and treat AI as a tool to supplement expertise and accelerate peer review, rather than as a replacement for reviewers (92). If reviewers use AI tools, they must verify the accuracy of the output (93), take accountability for the AI-generated review comments they submit, and ultimate decision-making accountability should always rest with experienced professionals (91,94). Errors in AI-generated content may undermine the quality, accountability, and fairness of academic evaluation. Human oversight is therefore always needed to prevent AI from amplifying existing biases in the literature or producing inaccurate peer-review outcomes due to hallucinations (95,96).

Sub-scenario 10. Using AI for decision-making among reviewers

Whether among those who embrace AI as an auxiliary tool integrated into the review process (97,98) or among traditional reviewers who insist that peer review must be performed independently by humans (62,66,99), the academic community has largely reached consensus that reviewers should not use AI for decision-making. For example, although we did not find specific rules on AI decision-making from institutions such as ICMJE and COPE or from publishers and journals, the small amount of available information prohibits AI decision-making: STM, ISTIC, and CMA all explicitly prohibit using AI for decisions. Most scholars believe that AI may amplify bias or even hallucinate, making it difficult to provide critical, constructive, and relatively fair review comments. Therefore, peer-review decisions should be made by human reviewers (95,100,101). Although AI can efficiently perform routine and standardized review tasks, including automated paper screening, flagging grammar problems and formatting details, helping reviewers avoid plagiarism risk, and highlighting similarities in sentences and data, it still has obvious shortcomings in core tasks involving deep scholarly judgment (88,102,103). In the core decision-making stage of novelty assessment, AI struggles to understand the context of highly specialized scientific topics, recognize the value of unconventional methods, evaluate complex results, figures and tables, or grasp broad meanings and subtle distinctions among words. As a result, it may overlook important considerations and fail to accurately judge the novelty and importance of content (38,94,104,105). AI is also insufficiently rigorous in methodological review, often focusing excessively on specific operational suggestions such as increasing sample size or adding experiments, while neglecting overall methodological rigor, and its criticisms often lack specificity (106,107). Its ability to review figures and tables is also poor (105). Kadi et al. used ChatGPT 4.0 to peer review 15 published short papers and found that AI had difficulty grasping whole-text logic, data, references, and methodological flaws, and could not make correct decisions (108). Marrella et al. found that peer review using ChatGPT-4o had an extremely high acceptance rate (mean 95%), far higher than the actual decision standards of human reviewers, with only approximately 30% agreement with decisions of high-impact journals and approximately 70% agreement with decisions of low-impact journals (109,110). Therefore, in current scholarly publishing, decision-making by human experts remains the unshakable cornerstone of peer review, and AI can only play an assisting role.

Sub-scenario 11. Using AI among reviewers to generate review comments or to assist/supplement review comments

In contrast to the highly consistent academic prohibition against AI making review decisions, attitudes toward reviewers using AI to generate or assist in generating review comments include not only prohibitive positions but also open or conditionally limited open positions among some institutions, publishers, and journals.

From the institution’s perspective, STM, ISTIC, and CMA have explicit rules stating that review can only be completed by humans and prohibiting the use of AI to generate review comments. However, because WAME’s recommendations state that reviewers should inform authors and editors if they use AI, it can be inferred indirectly that WAME permits reviewers to use AI to generate, assist with, or supplement review comments. No explicit rules permitting or prohibiting reviewers’ use of AI to generate or assist in proposing comments were found in the remaining institutions.

From the publisher’s perspective, Taylor & Francis explicitly prohibits using AI to analyze or summarize an article or any part of it. MDPI prohibits any AI use beyond permitted uses, including prohibiting AI from generating review comments or assisting in supplementing review comments. Additional searching found that Science also prohibits AI tools in peer review to prevent potential bias and misconduct (87). However, Frontiers states that reviewers may use publisher-provided internal AI tools to assist peer review, summarize manuscript content, highlight analytic priorities, and help draft feedback. Wiley also states that, without mentioning manuscript details, AI may be used to organize thoughts and process feedback, or reviewers may consult AI about general questions related to research methods or statistical methods. Two OUP journals also permit reviewers to use AI to refine their own comments. Further searching showed that these disagreements are very clear. One study found that among the peer-review policies of the top 50 dental journals in 2024, 38% completely prohibited such use, 10% did not include explicit rules, and 52% allowed GSM use under specific conditions (90).

In the literature, the vast majority of researchers still oppose reviewers directly using AI to generate complete review comments, while a minority recognize locally deployed AI review tools. Chauhan et al. proposed that, when human reviewers disagree, AI might be used to generate a third review comment for reference (111). However, AI not only has difficulty providing critical, constructive, and fair review comments (100), but may also cause reviewers to over rely on AI, reducing their engagement and critical-thinking ability; cause junior scholars to miss deep critical guidance from traditional peer review; and allow insufficiently scrutinized research to mislead investment of enterprise resources (92,112). In addition, AI-generated review comments pose technical and ethical risks. Generated content has obvious AI traces, and quality depends on the precision of instructions. Comments may be inconsistent or even randomly generated, and may also cause algorithmic bias and weaken interaction and collaboration between authors and reviewers (92,105,109,112). Nature’s large-scale survey of 5,000 scholars worldwide showed that 66% opposed using AI for initial peer review, 21% believed only locally deployed AI tools could be used, 8% believed AI could be used if AI use was disclosed, and only 5% believed AI could be used for review without disclosure (6).

Compared with AI directly generating complete review comments, the literature indicates higher acceptance among scholars for AI-assisted peer review, although clear disagreements remain. Opponents point out that AI-assisted review has many problems that cannot be ignored and may negatively affect the scientific rigor and fairness of review decisions. For example, Russo et al. found that scores from AI-assisted review were significantly higher than those from human review. The manuscript acceptance rate was 4.9% higher than that of human-reviewed papers, and the acceptance rate for borderline papers increased by 31.1%, increasing the risk that low-quality studies would be published (113). Supporters, however, believe that, as long as the major premises of ensuring manuscript confidentiality, disclosing AI use, having reviewers ultimately verify output comments, and holding reviewers responsible are met, AI can provide reviewers with assistance in many ways to improve review efficiency and quality (97,114-117). First, the literature indicates that AI can improve the efficiency of basic processes, including automatically checking manuscript grammar and plagiarism, identifying structural problems such as inconsistencies in terminology and references, helping organize review ideas, and checking manuscript formatting and other mechanical tasks (94,118). Second, AI can also expand the breadth of peer review by providing reviewers with interdisciplinary knowledge support, offering multidimensional review comments, serving as a benchmark for peer review to reduce extreme differences among comments, and assisting reviewers in evaluating the introduction and literature-review sections of manuscripts (61,109,111). Biswas et al. showed that AI’s critical analysis was consistent with that of human review experts and demonstrated excellent ability to identify methodological flaws, offer deep insights into theoretical frameworks, and evaluate the overall contribution of articles to their respective fields, thereby enhancing and supporting the work of review experts (105). Liang et al. also found that more than half of the comments produced by GPT-4 were consistent with those of human reviewers, and 82.4% of researchers believed that a small portion of GPT-4-generated review feedback was more helpful. Although AI has not yet reached a level regarded as “generally very useful”, it can already surpass some lower-quality human peer reviews (107). Khoshhal et al. further proposed that reviewers should be encouraged to become familiar with journal AI tools and provide feedback to editors on their effectiveness (119).

In summary, institutions, publishers, and journals have not reached consensus on reviewers’ use of AI to generate complete review comments or to assist in supplementing review comments, and clear disagreements remain. More scholars oppose reviewers’ use of AI to generate complete review comments, while a small number recognize locally deployed AI review tools. Disagreements also remain regarding the use of AI to assist in supplementing comments.

Sub-scenario 12. Using AI among reviewers to optimize the readability of review comments

Compared with reviewers’ use of AI for decision-making or for generating or assisting in supplementing review comments, the 6 organizations show somewhat higher acceptance of using AI to optimize the language/readability of human-generated review comments. ICMJE, WAME, and ISTIC all permit this. However, STM separately emphasizes that, although authors may use AI to optimize manuscript language, reviewers may not use AI to optimize their review comments, because STM considers confidentiality and privacy-protection requirements in peer review to be more important. This relates to the possibility that uploading review comments to an AI tool may leak important manuscript information because the comments are too detailed.

Similarly, publisher requirements are divided. Taylor & Francis, MDPI, and Wiley respectively allow reviewers to use AI to improve review language, check grammar, structure, spelling, punctuation, and formatting, and improve the clarity and tone of comments or translate comments. However, SAGE, OUP, and Elsevier completely prohibit this.

Attitudes in the literature are mostly supportive. Many researchers state that reviewers should be allowed to use AI tools for language polishing after completing their original review comments, provided that reviewers carefully verify the accuracy of the final output (77,82,90,98). Specifically, Biswas et al. believe that AI can help reviewers overcome language barriers, reduce the time required to write review comments, and improve the writing quality and standardization of review comments (105).

In brief, regarding whether reviewers may use AI to optimize the readability of review comments after independently completing the review, only the literature shows one-sided support, while clear disagreements exist among institutions and publishers.

Sub-scenario 13. Using AI among reviewers to check manuscript revisions and response letters

Corresponding requirements for this scenario currently contain substantial gaps. We did not find relevant rules in ICMJE, COPE, or ISTIC. Nor did we find specific publishers, journals, or studies separately mentioning whether reviewers may use AI to check manuscript revisions. However, some inferences can be made indirectly from broad AI-use requirements in few institutions. For example, WAME Recommendation 4 mentions that AI use in this process should be declared, which indirectly suggests that WAME permits it. In addition, because STM and CMA comprehensively prohibit reviewers from using AI during peer review and no separate exemption for this stage was found, it can be inferred indirectly that they also prohibit it.

Policies and requirements on use of AI among editors and publishers

Sub-scenario 14. Confidentiality and disclosure requirements

Regarding confidentiality, major institutions, publishers, and the literature show a high level of consensus similar to the requirements for authors and reviewers: editors and publishers also should not upload all or part of manuscripts, images, figures, or any communications related to manuscripts under review to any AI-containing system that cannot ensure confidentiality. This includes ICMJE, STM, WAME, MDPI, Elsevier, Frontiers, SAGE, OUP, and Taylor & Francis, among others, except for a small number of institutions and publishers that have not defined this issue. Discussions in medicine and the social sciences both point out that submitted manuscripts often contain patient privacy or clinical-trial data, while some cloud-based large language models may, under default settings, retain and reuse uploaded content for training, increasing the risk of leakage of sensitive manuscripts (60,88). Therefore, editors and publishers are advised to prioritize secure and reliable internal or controlled AI tools that have been approved by institutions and have audit-log functions; to carefully review terms of service and data policies before deployment; and to avoid processing full manuscripts or details identifying individuals or institutions in general cloud-based systems (60,89). Accordingly, both STM and ISTIC emphasize that publishers may use and are encouraged to proactively develop customized or locally deployed AI tools that comply with data-security and ethical standards, or to establish cooperative relationships with professional generative AI service providers that meet security, privacy, and confidentiality requirements, thereby preventing leakage risks at the source. Some scholars further argue that, in the long term, publishers developing or purchasing locally deployed AI tools for use by authors, reviewers, and editors under a unified security framework may be a feasible direction for balancing efficiency and confidentiality. This echoes the policy orientation of STM, ISTIC, and other institutions that encourage development of proprietary tools (120).

Regarding disclosure, although no rules on disclosure of AI use by editors and publishers were found in STM, ISTIC, or CMA, ICMJE, COPE, and WAME all emphasize that editors and publishers also need to clearly disclose to authors and reviewers any AI-related stage or AI automation used in the publication process, including the specific purpose, tool, application scenario, and method of use, such as manuscript preparation, screening, and copyediting. COPE further emphasizes that publishers should ensure transparent disclosure of the training process, algorithms, databases, known biases, and other limitations of the AI tools they use. However, among the 8 publishers and 139 journals, we did not find any rules on this issue. Only 2 OUP journals mentioned that, if editors use AI to refine their comments, this information must be disclosed to reviewers and article authors. This situation among publishers and journals differs sharply from the requirements of ICMJE, COPE, and WAME, even though many of these publishers and journals claim to follow the AI policies of ICMJE, COPE, or WAME. Scholars have focused criticism on this asymmetry, in which publishers and journals require only authors and reviewers to disclose AI use but do not disclose their own use, arguing that this practice weakens the fairness and credibility of AI-use standards. Relevant literature states that journals should not only require authors to report AI use in research and writing, but also clearly explain to authors and reviewers when and how natural language processing systems or other AI tools are introduced into editorial workflows, how their outputs are evaluated, and what investigation and appeal mechanisms will be used if misconduct or disputes related to AI use arise (91,94,121-124). The joint statement by editors of bioethics and humanities journals similarly emphasizes that any use of generative AI should be transparent to authors and should not become a hidden basis for review recommendations or editorial decisions (124). From a broader perspective, the “black box” nature of AI systems itself weakens the explainability of scholarly decisions. Therefore, whether AI is used for technical screening, review assistance, or content analysis, it should be disclosed to relevant parties so that authors and reviewers understand how AI participates in decision-making and, when necessary, can challenge decisions formed on that basis (94,121). Interestingly, some scholars also remind us from a historical perspective that current requirements for disclosing generative AI may be in a transitional period, just as almost no one today states in a manuscript that they used a particular word-processing software (125).

Therefore, there is strong consensus that publishers and editors need to ensure manuscript confidentiality. However, although major institutions and most scholars agree that publishers and editors also need to disclose in detail the various ways in which AI is used in the publication process, the vast majority of publishers and journals have not established equivalent self-requirements.

Sub-scenario 15. Verification and accountability

On this issue, institutions, publishers, and the literature agree that editors and publishers need mechanisms to ensure accuracy and effectiveness verification, and human accountability when using AI. For example, ICMJE, STM, COPE, and ISTIC all have explicit provisions. SAGE states that, when editors use AI to identify reviewers, reviewer identities must be verified in advance. OUP requires that editorial outputs reflect the editor’s own professional expertise and oversight. Elsevier states that editors bear full accountability for the review process, final decision, and communication of outcomes. The literature emphasizes that tool selection should be based on verifiable reliability and applicability rather than mere convenience or cost. Only systems that have undergone internal testing, peer evaluation, or third-party review and show stable performance are suitable for formal deployment, otherwise it is difficult to claim that AI use has been adequately verified (121). Accordingly, scholars require editors-in-chief and editorial teams to have a basic technical and functional understanding of the AI tools they use, to keep track of version updates and application changes, and to be able to critically interpret AI output, identify its limitations, and make independent judgments accordingly. In this respect, systematic AI training for editors should be regarded as a necessary component of professional development (60,121). A survey of biomedical journal editors-in-chief also showed that the vast majority of respondents believed it was highly necessary to provide editorial teams with training courses on the risks and responsible use of AI tools, so that they can independently verify AI suggestions and analyses while benefiting from improved efficiency (126). From ethical and legal perspectives, machines themselves cannot be accountable entities and cannot bear an obligation to explain themselves to the scholarly community or the public. Therefore, when erroneous publications or disputed manuscripts occur, accountability should remain with editors and publishing institutions (124,127). Multiple articles also argue that editors should be accountable for AI-generated content cited or adopted in any form. At the same time, editorial accountability for key decisions such as revision, acceptance, or rejection cannot be transferred. Even when such decisions refer to AI suggestions, they must be made ultimately by human editors, who are accountable for the outcomes (91,128-130).

Sub-scenario 16. Using AI among publishers and editors to check manuscript technical and compliance issues

Technical and compliance issues here refer to checks that do not involve substantive assessment of manuscript innovation, value, methodological design, statistics, or data. Instead, they include checks of manuscript scope fit, manuscript type, AI-generated content rate, ethics, research integrity, copyright, plagiarism, completeness of materials, and whether required content modules, such as conflict-of-interest statements, are missing.

ICMJE and ISTIC have policy gaps on this issue, while the remaining 4 institutions are evenly divided between opposition and support. On the one hand, STM states that editors and publishers should not use public platforms such as ChatGPT to handle such checks, because privacy, confidentiality, and copyright risks may exist. Similarly, CMA directly states that editors should not use generative AI for initial manuscript screening and standardization, without providing a detailed explanation of the scope of initial screening and standardization. On the other hand, COPE states that, except in situations that may raise concerns, retractions, institutional misconduct, or research-integrity evaluations in which editors cannot rely entirely on AI decisions, editors and publishers may use AI-driven automation technologies to improve processing speed, verify quality, assess standards, and advance the peer-review process, provided that AI does not directly decide acceptance or rejection. This may be necessary, for example, when AI detects that a manuscript lacks informed consent and automatically requests supplementation, or when AI identifies a human face and flags a possible privacy violation. WAME also mentions that editors need validated tools to help detect AI-generated or AI-modified content.

In contrast, although the vast majority of publishers and journals lack corresponding rules on this issue, the limited explicit rules and actual actions of publishers and journals express support. Elsevier and some sampled journals under Elsevier explicitly state that editors are allowed to use AI for technical checks of submitted manuscripts, such as matching journal scope, detecting duplicate submissions, checking compliance with submission guidelines and assessing completeness, or conducting research-integrity checks to ensure policy compliance. AI-driven pre-screening review tools have gradually become normalized in many journals. For example, plagiarism-detection software such as iThenticate, image-duplication and manipulation-identification systems such as Proofig, ImageRights, and TinEye, and AI-generated content detection tools such as GPTZero and the Scribbr detector have been integrated into submission and editorial workflow systems to identify obvious text and image misconduct risks before external review (131). Elsevier’s “Evaluate Manuscript” platform, based on Editorial Manager, automatically assesses a manuscript’s fit with the journal scope, potential citation impact, and similarity to existing literature by comprehensively analyzing the abstract, keywords, references, and publication record, helping editors identify manuscripts more deserving of external review within limited time (132).

The literature also shows a very positive attitude toward this issue. First, in manuscript screening and scope review, a study using medical journals as examples showed that, when appropriate prompting strategies were used, GPT-4 Turbo had high sensitivity and specificity in judging whether titles and abstracts matched journal scope (88). Second, in reviewing whether text is AI-generated, because traditional similarity detectors have already struggled to balance high sensitivity and specificity for detecting AI-generated text (3,48), and because humans, including editors, still have very limited and even near-random accuracy in identifying AI-generated content (48,133-135), reasonable use of AI detection tools with high sensitivity and specificity has become urgent. Editorial offices and publishers should neither rely entirely on a single detection tool (89) nor assume that experienced editors can easily identify AI text. Instead, they need to continuously update tools, provide editors with targeted training, and improve understanding of AI-writing characteristics and detection results (88,134). Third, in ethics review, studies have shown that natural language processing technology can automatically scan manuscripts for keywords and sentence patterns indicating informed consent, ethical approval, or data sharing, compare the extracted information with preset ethics-statement templates or databases, and quickly flag missing or ambiguous statements, thereby reducing post-publication disputes caused by incomplete ethics reporting (136-138). Fourth, in conflict-of-interest and authorship review, some work has attempted to use stylometry and clustering algorithms to identify potential authorship anomalies, ghostwriting, and abnormal collaborations by analyzing writing style, grammatical features, and textual organization (97,131), although this work remains exploratory. Fifth, in reference checking, reference-checking tools such as Scite can analyze how manuscripts use references and identify citation combinations that are inconsistent with the text’s argument or clearly inappropriate, providing editors with clues to detect citation-abnormality risks before external review (139).

In summary, regarding whether publishers and editors may use AI to check the technical and compliance aspects of manuscripts, many gaps remain in institutions, publishers, and journals, and existing institutional policies differ. Nevertheless, publishers, journals, and academia have expressed support through actual actions and publications.

Sub-scenario 17. Using AI among publishers and editors for content or substantive review and decision-making

As noted above, content review here refers to assessment involving manuscript innovation, value, methodological design, statistics, data, and other substantive content, while decision-making refers to decisions to accept, request revision, or reject a manuscript.

Compared with the technical and compliance checks discussed above, rules from institutions and publishers regarding whether editors may use AI for content review and decision-making also contain many gaps. Existing rules generally show consensus toward non-permission. COPE states that AI cannot make decisions alone and that editors must participate in decision-making. Frontiers, MDPI, and Elsevier also explicitly prohibit generative AI or AI-assisted technologies from participating in manuscript decisions, and Elsevier further states explicitly that editors are not allowed to use AI for manuscript review.

Regarding whether publishers and editors may use AI for decision-making, the literature is generally consistent with existing institutional and publisher views. The joint statement by editors of bioethics and humanities journals notes that generative AI should not be the sole basis for editorial decisions, and publishers should not replace human reviewers and editors with AI for the sake of efficiency. Human editors must retain dominant right and accountability for final decisions (124). Another article goes further, arguing that AI systems should not be used for key editorial decisions, such as topic review, whether to accept a manuscript, and whether to provide special promotion. It proposes incorporating relevant standards into “Good Publication Practice” and applying severe sanctions, such as removal from authoritative lists or citation bans, to journals that rely heavily on AI for decisions while lacking transparency and accountability mechanisms (99). This is because, if editors procedurally rely heavily on AI scores and recommendations, power and accountability become disconnected, authors have difficulty tracing the reasons behind decisions, and it becomes difficult to determine who should be accountable when errors or unfairness occur (53,97,99,124). However, AI can serve as a partial auxiliary reference for editorial decisions, and technical feasibility has also been demonstrated. Models trained on peer-review datasets such as PeerRead can use review texts to predict manuscript acceptance or rejection, with predictive accuracy between 74% and 86%, and show good consistency with editorial decisions in overall trends (140). A survey of biomedical journal editors-in-chief showed that most editors-in-chief rated their trust in AI-generated peer-review content at only about 6 out of 10, and only a small number of respondents trusted AI more than human reviewers. Overall attitudes tended to be cautious and reserved (126). These results empirically suggest that, at least at the current stage, editors-in-chief are more willing to regard AI as an auxiliary decision-support tool rather than an entity that can independently bear decision-making accountability. It is worth noting that the literature also contains radical imaginings of an “AI editor-in-chief”. One article envisioned allowing neural networks to fully take over editorial work, from screening manuscripts and selecting reviewers to making final decisions, arguing that AI would neither tire nor be influenced by personal bias or conflicts of interest and could theoretically achieve more efficient and fairer decisions (141). However, scholars generally acknowledge that machines cannot bear legal accountability for decisions, and the scenario in which editors-in-chief completely cede their role to AI is unrealistic (126).

However, regarding whether publishers and editors may use AI for manuscript content review, the literature expresses more positive signals of support. Such AI tools can provide editors with information on the preliminary quality of manuscripts through a relatively unified set of criteria, reducing the human workload to some extent (121,142). First, in methodological and statistical review, multiple tools have been used to help editors identify obvious flaws in study design and data analysis. Statcheck automatically compares statistics and P values in the text and appendices to check whether statistical reporting is internally consistent. StatReviewer further provides rule-based assessment of the choice and appropriateness of statistical methods, evaluates their consistency with the research question, and flags potential statistical misuse or missing key information (121,143). Scholars further suggest that editors upload manuscripts together with reporting guidelines such as STROBE, TRIPOD+AI, and TRIPOD-LMM to systems such as NotebookLM, and ask AI to check whether study design and reporting comply with the corresponding guidelines, reviewing items such as sample-size setting, variable control, and statistical methods, and generating targeted revision suggestions (144). Automated tools similar to OSF SCORE have also been envisioned for rapidly scoring the reproducibility of research conclusions and the strength of evidence, thereby flagging manuscripts that may have major methodological or statistical problems at an early stage and highlighting studies that perform well in design rigor and evidence quality (142). Second, in evaluating innovation and value, some studies have proposed that editors may refer to indicators such as innovation signals, citation trajectories, and “confidence scores” automatically generated by open-science platforms to make preliminary judgments about a manuscript’s future scholarly impact and the credibility of its conclusions (142,145). Third, in data review, machine-learning algorithms can flag possible serious data missingness, intentional underreporting, or data fabricated to produce desired results by identifying abnormal data or internal inconsistencies, thereby sending warning signals to editors (97). Although AI brings many conveniences, the literature also emphasizes that editors face corresponding challenges when using AI in this scenario, which places strict demands on human re-review and independent critical thinking. On the one hand, AI’s ability to fabricate may evade AI detection tools. One study reported that carefully refined LLM-generated synthetic medical datasets can contain substantially fewer detectable statistical anomalies, with some even passing forensic analysis as authentic, highlighting the limitations of relying solely on statistical patterns to identify fabricated data (146). Therefore, scholars remind editors and publishers that, while actively using AI to help identify abnormalities in data and images, they should not simply treat the output of a single algorithm as a conclusion that data have been fabricated. Instead, these signals should be judged comprehensively together with domain knowledge, research background, and other evidence, and deeper human review should be initiated when necessary (147). On the other hand, such AI review tools may harm academic fairness and diversity. Models trained on past publication and citation data are likely to inherit or even amplify existing disadvantages in the system against low-income countries, marginal fields, and non-mainstream methods, especially when such manuscripts are already underrepresented in training data (97). In addition, existing models have limited ability to understand ambiguous text, identify interdisciplinary innovation, and handle complex methodology. They are better at identifying manuscripts that conform to existing highly cited patterns and often give conservative or even negative assessments of truly innovative manuscripts that depart from conventional paradigms (53,97).

Taken together, institutions, publishers, journals, and the literature have reached broad negative consensus regarding editors’ use of AI for decision-making. However, disagreements remain regarding AI use for manuscript content-depth review, reflected in negative attitudes among institutions, publishers, and journals and positive attitudes in academia under the premise of paying attention to risks and maintaining independent critical thinking.

Sub-scenario 18. Using AI among publishers and editors to select reviewers

On this issue, we found no relevant provisions among the included 6 institutions, 6 publishers, and 139 journals. However, from publicly available rules and real-world practice, academia generally appears to permit such use. For example, both SAGE and Elsevier state that editors may use AI to search for and identify relevant reviewers. In addition, many commercial publishers have integrated AI-driven reviewer recommendation systems into editorial platforms. Clarivate’s Reviewer Locator (148), Elsevier’s Find Reviewers (149), and Reviewer Discovery developed by Aries in cooperation with ProQuest (150) are typical examples. Data show that, in 2024, approximately 60% of Elsevier journal editors used Find Reviewers to assist in identifying reviewers, up from approximately 40% two years earlier, reflecting increasing editorial acceptance of such tools (151). These tools use databases such as Web of Science, Scopus, and ProQuest to automatically generate lists of candidate reviewers related to manuscript topics, which editors then manually screen and confirm.

In actual operation, studies particularly emphasize several points that editors should note. First, AI reviewer recommendation may reduce or amplify bias. With the help of large language models, editors can identify scholars from non-native English-speaking countries or scholars who are less well known but have some accumulated expertise in specific subfields. This helps expand reviewer pools, increase diversity in reviewer teams, and reduce biases related to reputation or personal networks. Editors generally recognize the value of AI in identifying potential reviewers beyond their usual field of attention (152-154). However, most existing commercial tools mainly rely on large databases such as Web of Science or Scopus, together with the opacity of LLM training data, which may unconsciously embed preferences for English-language literature, mainstream disciplines, and institutions into the reviewer-recommendation process, thereby amplifying existing inequalities during selection (155). Another randomized trial compared reviewer lists recommended by GPT-4o and human editors and found that the model, under default prompts, clearly favored men and scholars from high-income countries (152). This is consistent with findings that human editors’ selection of peer-review lists contains gender bias (156). Only when requirements for “gender and geographic diversity” were explicitly added to the prompt could GPT-4o generate lists that were broadly balanced by gender and more geographically diverse (152). As early as 2017, the 2030 peer-review outlook report jointly released by BioMed Central and Digital Science anticipated that AI could identify new potential reviewers from sources that editors might not have considered, and that one core element of future reform would be finding and innovating methods to identify, verify, and invite peer reviewers, using technology to more closely match reviewer expertise with manuscript content (157). Second, AI-recommended reviewers may involve hallucinations. Lists recommended by GPT-4o were found to contain a small number of names that could not be traced or belonged to deceased scholars, suggesting that large models may “hallucinate” nonexistent reviewer identities (152). In addition, approximately 10% of AI-recommended lists were judged by editors to be fabricated or unusable (154). Third, editors need to review AI-recommended reviewers. Scholars currently tend to position AI in reviewer selection as a supervised support tool and argue that AI recommendations should not be placed directly on the final list for automatic invitations. Editors need to conduct rechecks (88,111,155). More studies may be conducted in the future to evaluate the effectiveness of reviewer lists recommended by AI, such as whether the review quality of AI-recommended reviewers is comparable to that of traditionally human-selected reviewers.

Sub-scenario 19. Using AI among publisher’ and editors to summarize manuscript content and draft/improve comments

Rules from the 6 institutions and the vast majority of publishers and journals are still absent for this issue, while existing policies and attitudes in the literature tend to support it. Specifically, Frontiers states that editors may use internal AI tools to summarize manuscript content, highlight key analytic points, and help draft constructive feedback. Wiley also believes that editors may use AI to organize thoughts and prioritize feedback. Two OUP journals also state that editors may use AI to refine their own comments, but disclosure is required. Scholars believe that such tools can automatically generate summaries and synthesize key points, enabling editors to form a relatively clear overview of complex manuscripts and multiple comments within a short period and providing an information basis for later writing decision letters and refining feedback. For example, the UNSILO system can automatically extract key concepts and topics from manuscripts, providing editors with a quick way to understand the research focus and main contributions (143). Consistent with this, large language model-based tools such as SciSummary and Scholarcy can generate structured summaries that clearly present study objectives, methods, main results, and conclusions. They are especially helpful for forming an initial judgment of the overall framework of articles in a short time when handling interdisciplinary manuscripts or manuscripts beyond an editor’s core expertise (121). Some editorial platforms have already embedded proprietary summarization engines into workflows to automatically generate “editorial summaries”, with the aim of improving internal work efficiency while enhancing consistency and transparency in decision communication (121,158). The PeerRead model can extract key information from multiple review reports and generate review-style text, helping editors quickly grasp the overall review tendency and main points of dispute (140).

Sub-scenario 20. Using AI among publishers and editors for language polishing and copyediting

Similarly, many institutions, publishers, and journals have no rules on this issue, while existing policies and attitudes in the literature are mostly supportive. Specifically, COPE mentions that publishers and editors may use AI-assisted tools to edit manuscript, but should avoid changing authors’ position or tone as much as possible. Wiley explicitly states that editors may use AI to improve the clarity and tone of comments or translate comments. Elsevier also supports editors using AI for proofreading, copyediting, and identifying inconsistencies or errors in final articles. One survey showed that approximately 70.8% of editors-in-chief recognized the advantages of AI in language and grammar support, and nearly 60% agreed that AI can simplify technical checks and formatting. Language polishing and proofreading were regarded as among the application scenarios most suitable for introducing AI (159). Scholars also generally believe that generative AI, as an auxiliary tool for editors and publishers, can substantially reduce workload in grammar, spelling, punctuation, and basic sentence fluency, especially for resource-limited journals and non-English contexts (158,160,161). For example, Wordvice AI achieved an average correction rate for grammar, spelling, and punctuation errors of approximately 77% of that of human editors, and overall outperformed general writing software such as Google Docs and Microsoft Word, especially in highly technical and structured academic texts (162). ChatGPT can effectively identify and correct spelling, punctuation, and capitalization errors and has potential value for editors (10,163,164). Some scholars therefore suggest that editors can use custom instructions and prompts to apply ChatGPT to improve sentence clarity and paragraph coherence and to support non-native English-speaking editors (165). However, although multiple studies have recognized the application potential of AI in editorial language polishing and copyediting, there is still considerable room for editors to adopt it in actual work. A survey of 59 medical journal editors-in-chief found that the proportion of editors using AI for substantive language review in their work is currently only about 9% (126). This may be related to lagging publisher and journal policies and training.

Sub-scenario 21. Using AI among publishers and editors for formatting optimization and typesetting

We found no relevant rules for this application scenario among the 6 institutions, 8 publishers, and 139 journals. Similarly, the literature contains little discussion of editors’ and publishers’ use of AI in formatting optimization and typesetting. Unlike language polishing, formatting optimization almost does not involve substantive intervention in textual content, and therefore in theory carries lower risk. Academic acceptance of AI applications in this scenario may be higher. One survey provides indirect support: nearly 60% of editors-in-chief agreed that AI can simplify technical checks and formatting (159), and formatting optimization and language proofreading both fall within AI application scenarios recognized by editors. At present, formatting optimization and typesetting are more often regarded as later stages of manuscript processing. The policy and practice gap in this area partly reflects the fact that current AI discussions remain focused on the front end of manuscript processing, with insufficient attention to back-end publishing workflows.

Policies and requirements on AI use among readers and in scholarly dissemination

Sub-scenario 22. Confidentiality and disclosure, verification, and accountability

Policies on this issue remain largely absent across the sampled institutions and publishers. The literature also did not identify relevant rules requiring confidentiality and disclosure when readers use AI. We found only that Wiley mentions that, if AI is used during scholarly dissemination, users should review the publishing agreement to confirm whether the work may be adapted into other forms, such as podcasts, blogs, or social media posts. Unless the agreement explicitly permits this, readers should not upload full articles or large amounts of copyright-protected text to AI or other platforms. Another article cautions that, although AI can facilitate the literature-reading experience, it carries risks such as hallucination, misinterpretation, and unstable accuracy of generated content. Readers therefore need the ability to distinguish and verify the truthfulness of AI-generated content to prevent the spread of false scholarly information (166,167).

Sub-scenario 23. Using AI among publishers or media to generate downstream content for social media, press releases, visualizations, and similar outputs

Similarly, policies from the vast majority of institutions and publishers are mostly absent on this issue. We found no requirements from STM, COPE, WAME, ISTIC, or CMA. Only ICMJE’s mention that “journals may apply AI and other tools to generate ancillary content, such as blogs and plain-language summaries, and editors need to ensure the accuracy and validity of such content” allows an indirect inference that ICMJE permits this use scenario. Among the 8 publishers and 139 journals, only Wiley was found to mention permission for use in such scenarios, including using AI to generate cross-platform content such as social media posts, newsletters, and presentation summaries.

Correspondingly, the literature has widely discussed how AI empowers the dissemination chain in medical publishing. Most scholars believe that AI can assist editors in content transformation, special-topic planning, data visualization, video production, online lectures, community promotion, and international dissemination, and can use social media data analysis to monitor and evaluate journal dissemination effects in real time (168,169), as well as the influence of AI-generated content technology on scientific and technical journal publishing and its responses (170). In addition, by analyzing reading data such as readers’ browsing frequency, dwell time, related comments, and links, AI can precisely push papers to readers and improve dissemination efficiency (168,169). However, some scholars believe that publishers’ use of AI to collect and analyze readers’ reading data and provide personalized recommendations during scholarly dissemination may carry risks of infringing readers’ personal privacy (171). In addition, while efficient and convenient intelligent generation empowers scholarly publishing, it also brings a series of publication ethics crises. Some scholars worry that false and misleading AI-generated content may damage journal credibility, and that problems may include homogenization, amplification of inherent biases, filter bubbles, and unclear copyright ownership (53,100,168,172). Moreover, AI may tend to favor hot topics and commercial value while neglecting public health needs, and it may be misused to manipulate social media exposure, citation counts, and altmetrics, distorting the authenticity of scholarly impact (53,168).

In this scenario, the few institutional and publisher attitudes available are supportive, while most policies are absent. Scholars also tend to support such use, while concerns coexist regarding privacy infringement, content homogenization, intensified bias, and filter bubbles.

Sub-scenario 24. Using AI among readers to extract key points and directly translate articles during reading

We did not find any information in the requirements of several institutions on whether authors are permitted, encouraged, or cautioned to use AI to extract article highlights or directly translate articles. However, STM emphasizes another point: readers are prohibited from uploading published manuscripts to public generative AI platforms because this may violate copyright, confidentiality, and privacy requirements, especially when the manuscript is not open access. Therefore, it can be inferred that, under STM rules, if authors upload the full text of a non-open-access manuscript to an AI platform for one-click translation, there is indeed a risk that the third-party AI platform will retain large amounts of full-text manuscript content that infringes copyright.

Similarly, the vast majority of publishers and journals have no rules on this issue. We found only that Wiley permits the use of AI to convert technical or complex language into expressions suitable for different groups, but users must check accuracy and tone because AI-generated content may oversimplify or distort research findings, or use a tone that does not fit the target audience or personal communication style.

The literature highly recognizes AI performance in this context, mainly by acknowledging that AI can assist readers in reading through multiple methods, including translating papers, extracting key points, and generating reading reports, thereby reducing reading difficulty and saving reading time. For example, the Sciengine platform is equipped with an integrated automatic translation tool that can translate Chinese journal articles into English, facilitating reading by international readers (83). ChatGPT can rewrite scientific articles into understandable, concise, and accurate text, present original data in a concise format, and substantially lower the reading threshold, making it applicable to patient education (173). Matalon et al. further found no significant difference in readability or content quality between AI-generated summaries and human-generated summaries, and that AI systems and reader digests can save physicians time (174). However, some researchers believe that AI may produce errors, accelerate the spread of misinformation, or reinforce existing biases and create filter bubbles (53,172). Therefore, readers still need to use AI-assisted reading critically and cautiously.

In summary, the few institutional and publisher attitudes available provide conditional support, while most policies are absent. Published literature tends to support AI use in scholarly dissemination and reader-facing applications, although concerns coexist regarding content errors, misinformation, privacy, copyright, content homogenization, bias, and filter bubbles.

Sub-scenario 25. Using AI among scholars to generate conference posters and PowerPoint presentations

It is very common for scholars to share their own or others’ scholarly outputs through conference posters or presentation slides. With the emergence of many so-called one-click AI tools for generating presentations, more and more people in the real world directly use AI to generate such presentations. Unfortunately, we found no relevant attitudes on this issue among the 6 organizations, 8 publishers, 139 journals, or in the literature. However, one study stated that AI can enable real-time transformation, tracking, and interpretation of research outputs after publication, helping authors transform research findings into engaging teaching cases, learning applications, or even games, thereby making educational dissemination more vivid (53).

Discussion

As shown above, across the life cycle of scholarly publishing, institutions, publishers, journals, and academia have reached consensus in some scenarios, while many scenarios contain disagreements or even policy gaps (Figures 2,3). Scenarios with disagreements and gaps need to be addressed as priorities in future policy updates. The discussion of these issues has been provided in the preceding sections and is not repeated here.

Figure 2.

Overview of consensus in attitudes toward AI use among academic institution, publishers/journal, and scholar throughout the scholarly publishing process. AI, artificial intelligence.

Figure 3.

Attitudes of institution, publisher, and scholar toward AI use by author, reviewer, editors/publisher, reader, and in dissemination. AI, artificial intelligence.

It is especially important to note that institutional AI policies have undergone multiple updates, and AI policies from institutions such as ICMJE are important points of reference for publishers and journals. However, many institutions label the AI policies they issue more as “Recommendations”, and multiple institutions explicitly emphasize in their policy documents that specific use scenarios should be governed by publisher and journal policies and editorial judgment, as seen with ISTIC and STM. Against this background, publisher and journal rules become extremely important. However, we found that many publishers do not define permitted uses clearly enough. For example, some publishers simply state that authors are permitted to use AI for literature reviews, but the concept of a literature review is extremely broad, including generating review content, generating review’s references, and generating review frameworks. Using only the words “literature review” is too vague. In addition, each publisher emphasizes different aspects when defining AI use. For example, SAGE and Frontiers place more emphasis on AI use in figures, data analysis, and visualization, whereas Springer Nature and Elsevier focus more on AI applications in images and videos and provide fewer rules for other aspects. All of the above further increases the difficulty for authors and reviewers in using AI in a more standardized manner.

In addition, many journals have no AI policies or introductions on their homepages, including in author instructions and other locations. Although many journals now have AI rules, there is still room to improve visibility. Regardless of where the journal presents these rules, whether the journal presents all AI requirements in full or briefly introduces the requirements and then provides a hyperlink to the publisher’s general AI policy, some journal AI rules are difficult for readers, authors, or reviewers to access. For example, some journals’ author instructions do not contain the term AI. Only the ethics section provides a hyperlink, and the linked page also contains no prominent mention of AI but instead redirects again by hyperlink to the detailed publisher AI policy. Given that current AI policies for many specific scenarios already contain disagreements and gaps, the above situations further increase the difficulty for readers, authors, and reviewers to know a journal’s AI policy before submission. Koulaouzidis et al. advocate that journals may consider adopting a tiered disclosure policy: Level 1, adding a summary statement on the first page of the author guidelines indicating whether specific AI rules exist; Level 2, providing a link to detailed guidelines, including examples of acceptable and unacceptable use; and Level 3, requiring a checklist form at submission to disclose any AI use (175), thereby ensuring disclosure clarity and consistency.

Furthermore, we found that, in policies regulating AI use across various scenarios, publisher and journal rules are far thinner and often absent for editors and publishers themselves than for authors and reviewers, even though many institutional policies and scholarly views emphasize that publishers should be subject to equivalent constraints. In the future, publishers and journals need to strengthen this area of policy to enhance scholars’ confidence in fairness and their support for other policies targeting authors and reviewers. Such details include whether the tool name, version, stage of use, and method of use should be disclosed, and disclosure locations such as decision letters, journal webpages, or article footnotes. More importantly, corresponding mechanisms, such as various forms of journal evaluation, need to be established to provide strong oversight of disclosure implementation by editors and publishers, as well as the completeness and truthfulness of their disclosures.

Moreover, although the academic community has currently reached consensus that “manuscripts should not be uploaded to third-party AI platforms”, relevant stakeholders have enormous demand for AI use across scenarios, creating a current contradiction. From 2016 to 2022, the global number of papers increased from approximately 1.92 million per year to 2.82 million per year, an increase of 47% (176). The continuously growing manuscript volume has increased review pressure, and overburdened scientists may turn to increasingly powerful large language models for peer review (177,178). Reviewers’ dependence on AI tools is gradually increasing, and undisclosed use of AI in peer review is also becoming more common (88,92,179). A submission review survey from the International Conference on Learning Representations showed that at least 15.8% of reviews were written with AI assistance (113). Peer-review texts from several top AI conferences, including NeurIPS 2023, CoRL 2023, and EMNLP 2023, also showed that 6.5% to 16.9% of content may have been substantially modified by large language models, exceeding simple spell checking or minor writing supports (180). If publishers and journals can provide more journal-level or publisher-level internal AI platforms in the future that ensure manuscript confidentiality and have undergone quality review, this will likely greatly reduce the workload of reviewers and editors, and may also become an important future solution for ensuring data privacy and data security. However, secure deployment alone does not establish an AI system’s suitability for substantive scholarly judgment. Any future expansion into such tasks should involve training and evaluation signals beyond historical acceptance decisions, potentially incorporating initially rejected research subsequently supported by independent validation. Outputs should provide calibrated uncertainty estimates and traceable, evidence-linked justifications that enable audit and challenge, rather than binary verdicts alone. Before AI assessments are assigned greater decision-making weight, prospective studies should compare AI and human assessments against longer-term research outcomes, including reproducibility and scientific contribution, without treating citation counts alone as evidence of quality. In parallel, platform governance must address privacy and intellectual property protection. Scholars have also proposed that AI platforms need to implement privacy-protection policies, and corresponding laws and regulations must keep pace with AI development to ensure that intellectual property rights are protected (181). The rapid development and use of AI are inevitable. Rather than asking users to exhaust themselves trying to protect their privacy, developers should prevent intellectual property infringement at the source, just as journals protect the intellectual property of all publications under their ownership from infringement.

Finally, in the AI era, in addition to formulating detailed policies for authors, reviewers, publishers, editors, readers, and scholarly dissemination, publishers and journals also need to continuously provide more free and publicly available learning and training resources to help improve the abilities of authors and reviewers. These include distinguishing legitimate research from fraudulent research, developing critical ability to distinguish AI-generated content from human-written text, and understanding the limitations and risks of AI-tool use (93,113,165,182). In addition, good feedback channels should be established to encourage users to become familiar with the AI tools adopted by their journals and to provide feedback to editorial offices on their effectiveness (119).

Strengths and limitations

Although previous reviews have summarized the general status of AI policies in scholarly publishing, to the best of our knowledge, our review provides a broader overview of this field by including the latest policies from 6 major institutions, from 8 impactful publishers and 139 journals from all quartiles under those publishers, and scholars’ view reflected in publications in multiple databases. Second, we divided the various scenarios into 25 detailed sub-processes/scenarios and, on that basis, integrated policy information on whether consensus has been formed, whether disagreements exist, and whether gaps remain. This provides highly valuable information for further detailed policy improvement, including a more detailed scenario framework for future organizational updates of AI policies and an important basis for publishers and journals to formulate future policies. Third, we conducted an in-depth discussion and analysis of the shortcomings of current AI policies and provided corresponding suggestions, which may support the implementation and design of clearer, more transparent, and more executable AI policies in the future.

At the same time, this review has its own limitations. First, given the large volume of information in this field, we conducted a narrative rather than a scoping review. We also did not perform a formal methodological quality or risk-of-bias assessment of the heterogeneous publications included. Conflicting findings and viewpoints were described and compared narratively, the absence of formal appraisal limits our ability to assess the relative strength of evidence across sources. Second, the included policy analysis may not represent policy positions of other organizations, publishers, and journals. Third, the literature search was confined to sources published in Chinese and English, within defined organizations, publishers and journals. Relevant studies and perspectives published in other languages, organizations, publishers, and journals may have been overlooked. Finally, this review reflects the policy and literature landscape at the respective search dates, subsequent policy updates and new publications may limit the currency of its findings.

Conclusions

In summary, across the various scenarios in medical scholarly publishing involving authors, reviewers, publishers and editors, readers, and scholarly dissemination, some consensus has been reached among policies from institutions, publishers, and journals and views from scholars, but many gaps and disagreements remain. Future policy updates should not only focus on addressing these disagreements and gaps, but also strengthen the specificity of policy content and avoid wording that is too general and insufficiently specific. In the public presentation of policies, user accessibility should also be considered. Repeated deep hyperlink redirections should be avoided as much as possible; complete policy text is highly recommended to go with visual summary to facilitate users’ access and understanding. In addition, policy requirements and public disclosure for publishers and editors themselves need to be strengthened. Finally, AI platforms that can ensure confidentiality and reliability should be developed, free and publicly available learning and training resources on AI use should be provided for scholars, and channels should be opened to encourage and accept feedback from scholars.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.