The Presence and Nature of AI-Use Disclosure Statements in Medical Education Journals: A Bibliometric Study

Background

AI has the potential to transform research practices by enhancing productivity in both research and writing. At the same time, its use introduces significant ethical and methodological challenges. When applied uncritically or without transparent reporting, AI may compromise research integrity through issues such as unclear attribution of intellectual contribution or obscured methodological decisions [1, 2]. Consequently, disclosure policies from organizations such as the International Committee of Medical Journal Editors [3], publishers and journals [4, 5] share a general emphasis on transparency and accountability. While specific requirements vary, authors are required to explicitly disclose whether they used AI-assisted technologies in the production of submitted work, and to take responsibility for all AI-produced material [6, 7].

However, emerging research suggests that AI disclosure may be lagging behind AI use. A 2025 Nature survey of 5000 researchers found diverging views of disclosure [8]. For instance, when asked whether they had used AI to write a section of a paper and not disclosed the AI use, 17% of mid-career researchers said ‘Yes’, and 47% said ‘No, but I would be willing to’. Database research also suggests that AI use outstrips disclosure. For instance, one comparative study of >5 million articles in the Dimensions database reported a 468% increase in clusters of ChatGPT-preferred positive terms (e.g., meticulous, intricate, commendable) in texts published in 2023 after the model became available; only 0.1% of those papers contained language suggesting disclosure [9]. Similarly, a 2024 analysis of abstracts published in high-impact orthopaedic journals found that of 28 containing AI-generated text, only 1 disclosed AI use abuzz with keynotes, panels, workshops, and papers related to AI [10]. But while it is increasingly apparent that scholars are using AI [11], the nature and pattern of our disclosure practices are less clear. In a recent interview study, medical education journal editors described infrequent experience with AI-use disclosures in submitted manuscripts [12]. Most suspected that medical education researchers might be using AI without disclosing, and worried that this may arise from a lack of clarity regarding when disclosure is necessary and what details are required.

To deepen understanding of AI-use disclosure in medical education, we require objective evidence of researchers’ current disclosure practices. Therefore, this study asks: What is the rate and nature of AI-use disclosure in empirical research papers in medical education journals?

Methods

We conducted a bibliometric study to identify and characterize AI disclosure statements appearing in medical education research articles. Our sampling frame included all articles published between January 1 and June 30, 2025, in the 24 leading medical education journals listed in the Medical Education Journals List, which is a set of journals created via bibliometric methods to identify core medical education journals that serve as central outlets for scholarship in the field [13]. We selected a 6-month sampling period to balance feasibility with adequate sample size, recognizing that this timeframe would preclude analysis of temporal trends but provide a robust snapshot of current disclosure practices. We excluded non-empirical publications (e.g., commentaries, letters, perspectives) both because we perceived the stakes to be higher for disclosing AI use in research than in commentaries and we anticipated that the conventional genre of empirical articles would support meaningful comparisons regarding disclosure content and location.

Screening and article selection were managed in Covidence, a web-based knowledge synthesis software. MA, HA, JC, and LM independently screened the titles and abstracts of all articles to exclude non-empirical publications. Full-texts of the remaining articles were then examined by MA and HA for AI-use disclosure statements. We classified articles as having an AI statement if they explicitly disclosed the use of an AI tool (e.g., ChatGPT) or described how AI was used in the preparation of the work. Articles that merely focused on, evaluated, or discussed an AI tool, without indicating that AI was used to generate or assist with the manuscript, were not counted as having an AI disclosure statement. For example, an article studying the effectiveness of AI in clinical reasoning, but lacking an explicit disclosure of AI involvement in the authorship process, would not be considered to have a statement. Articles with an AI-use disclosure statement proceeded to data extraction (Figure 1).

Figure 1

MA and HA each individually extracted data from approximately half of the articles using a structured Google Sheets template. They met regularly to ensure consistent interpretation, discussing ambiguous or complex disclosure statements and reaching consensus through discussion. For each article with an AI-use disclosure statement, we extracted:

• Bibliographic variables: DOI, article title, journal name, publication date, journal impact factor, publisher.

• Author-level variables (first author only): name, total publications, total citations, H-index, year of first publication, institutional affiliation, career position, and country of affiliation. These data were collected using Web of Science (WoS). Although we recognize that authors may have publications not indexed in the WoS, we viewed this as a reasonable proxy measure for publishing experience. We focused on first authors as they typically lead writing and analysis procedures. While we recognize that corresponding authors may also influence AI-disclosure, the first author focus ensured feasible study scope.

For bibliographic and author-level variables, in July 2025, we downloaded the metadata from WoS; if unavailable on WoS, we queried Scopus, except for author career position. Author career position was extracted from the article’s author information, or if unavailable, from ResearchGate. Additionally, based on a review of journal and publisher websites, we identified if AI reporting guidance was provided at the time of our study.

Disclosure-specific extractions focused on the name of AI tool(s) used, purpose of AI use, location of the disclosure statement (e.g., methods, acknowledgements, disclosures), verbatim text of the statement, statement word count, and the presence or absence of an attestation of responsibility (e.g., authors remain accountable for the content). We also extracted measures of transparency to situate AI disclosure statements within the broader context of established transparency practices. These included open access status, presence of a funding statement, conflict of interest disclosure, and data availability statement.

For analysis, we conducted descriptive statistics to characterize the bibliographic and author-level variables. To describe the content of AI-use disclosure statements, we conducted a descriptive content analysis [14] for AI tool used, the nature of usage, and any reference to responsibility attestation. Two researchers were involved in the analysis: each analyzed a subset of the disclosures, coming together regularly to discuss and resolve discrepancies.

Results

Over the study period, 2,762 articles were published in the medical education journals sampled. Of these, we excluded 716 (25.9%) non-empirical articles. Of the remaining 2,046 articles, 2.5% (n = 51) had an AI disclosure statement (Appendix A) [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65].

Thirteen of the journals included articles with disclosures. Of these, 12 (92.3%) provided author guidance on disclosure on their website/author instructions, or on their publisher’s website (Table 1). BMC Medical Education published nearly half the identified statements (n = 24, 47.1%), followed by Medical Teacher (n = 7, 13.7%) and the Journal of Surgical Education (n = 4, 7.8%). Nine journals did not include any articles with disclosures. Two journals (African Journal of Health Professional Education and BMJ Stimulation & Technology Enhanced Learning) did not publish any empirical research during the study period and, therefore, were not examined for AI-use disclosures.

Focus on Health Professional Education had the highest proportion of disclosure statements (25%), although this reflects a small denominator of total articles published, with 2 of 8 containing disclosure statements. Across the other 23 journals, the prevalence of AI-use disclosure statements was low, with no journal exceeding 7% of published articles (Table 1).

Discussion

Our findings highlight three primary issues for consideration in our field: the low rate of disclosure, the geographical pattern of disclosures concentrated in non-native English-speaking regions, and the practice of superficial, safe disclosures.

Our analysis suggests that AI-use disclosure is rare in medical education: only 2.5% of empirical articles published in 24 medical education journals in the first half of 2025 included a disclosure. This corroborates medical education journal editors’ subjective experience of infrequently seeing disclosures in submitted manuscripts [66]. It also resembles bibliometric results from other research domains: e.g., an analysis of academic radiology articles in 2024 found that 34 of 1998 manuscripts (1.7%) disclosed AI use [67]. The low rate could be interpreted in many ways. It may reflect low AI use. It may signal lack of awareness of AI use, such as when researchers use a tool not knowing that it is powered by AI. Or it may reflect a tendency towards nondisclosure of AI use. Some evidence suggests the last [68]: for instance, since 2023, analyses by “integrity specialists” have flagged hundreds of published research papers with obvious signs of undisclosed AI use, such as those that contain the phrase “regenerate response” [69]. Given that the field of medical education research is not immune to undisclosed AI use [70], we need to explore the underpinning influences on our field’s current rate of AI-use disclosure.

One influence is the lack of clarity. Not only are journal policies on AI use and disclosure shifting quickly due to the fast-changing AI landscape, journals’ positions vary. For authors submitting, and resubmitting, their manuscripts, disclosure requirements may appear murky and complicated. Another influence is the divergent attitudes among researchers about AI use and disclosure: surveys have clearly shown that we don’t agree about what AI should be used for and when those uses require disclosure [8, 11]. A third influence is psychological/cultural and institutional barriers that make researchers hesitant to disclose AI use to improve their writing. They may experience a sense of “demonization” of AI [71], inducing guilt or shame about using it to support their work [72] and fear of consequences such as academic stigma and evaluation bias [73]. These concerns have recently also surfaced in relation to research teams [74], where decisions about AI use and disclosure may be implicit, unevenly negotiated, or avoided altogether. Such dynamics can negatively impact team functioning, strain trust, and complicate authorship negotiations. Finally, the lack of published disclosures may act like a self-fulfilling prophecy: if we suspect others are using AI but we rarely see AI-use disclosures, that can send a tacit message that disclosure is unnecessary.

Based on frequency, our findings also demonstrate a particular geographical pattern to AI-use disclosure in the analyzed articles, with most authored by individuals from non-native English-speaking regions of the world. This likely explains the dominant use of AI tools for language-related tasks, like editing, translating and grammar checking, all of which are valuable to authors writing in English as a non-native language, and resonates with other studies that have reported higher AI use for writing support by non-native English researchers [75]. Given this finding, medical education needs to be aware of the possibility of stigmatization, given that non-native English academic writing is more likely to raise reviewers’ suspicions about AI use [76], and is also known to be prone to misclassification by GPT detectors [77]. The four negative disclosures in our dataset all came from authors in non-native English-speaking regions, which may suggest an attempt to deflect such potential stigmatization. Medical educators need to be aware that, rather than levelling the global playing field in academic publishing, AI-use disclosure might intensify inequity and bias against authors from some parts of the world [78], further exacerbating the global North dominance in the field’s literature [79, 80]. As a recent AMEE Guide cautioned, the equity implications of AI-use disclosure practices and policies require ongoing attention [81].

The disclosures we analyzed were almost uniformly brief, describing general uses such as “editing”, and lacking in detail that would allow a reader to understand precisely how the author engaged with the AI, iteratively refined and structured its work, and verified its outputs. We interpret this as “safe” disclosure practice: authors are mostly including the required elements of tool, task, and attestation, but rarely disclosing substantive, intellectual tasks, or elaborating the ‘how’ of their interaction with the AI, including specific verification practices underpinning their attestations of responsibility. On the one hand, this might be good news: the disclosed uses we found mostly map onto the “acceptable” uses outlined in a recent framework distinguishing ethically acceptable, contingent and suspect uses of AI [82]. On the other hand, though, we worry that our dataset includes few disclosures in the framework’s “contingent” or “suspect” ranges, not because researchers are not using AI in these ways, but because they perceive themselves to be on shaky ground in these areas. While our study was not designed to judge the accuracy of the AI-use disclosures we analyzed, the predominance of “safe disclosures” suggests that future research should explore how authors are deciding what and how to disclose.

The combination of low disclosure rates and safe disclosure practice produces a “transparency paradox” [73], in which mandatory disclosure that does not attend to social complexities leads to both nondisclosed AI use, and disclosure “theatre” in which published disclosures are more performative than informative [83]. One possible solution is for journals to create dropdown menus of possible AI uses and verification strategies, signaling sanctioned uses and expected levels of detail. These menus might draw from emerging disclosure frameworks specifying tiers of usage [84] or using CREDiT-like author contribution frameworks [85]. Another solution might be to shift away from ‘disclosure’ language altogether and make AI use a routine part of methods reporting, as recommended in the recent AMEE Guide [86]. This could help to normalize writing about AI use in manuscripts, thus reducing the negative valence associated with disclosure in the current literature. Given our results, however, we anticipate that authors will need explicit guidance to include responsibility attestations when reporting AI use within their Methods.

Conclusion

AI-use disclosures in medical education journals are rare and appear mostly in work from non-native English-speaking regions of the world. A shared disclosure practice is evident: name the tool and affirm author responsibility, but describe use superficially. This suggests a practice of “safe” disclosure that may fail to satisfy the goal of ensuring transparent and ethical AI use in research.

Additional File

The additional file for this article can be found as follows:

Ethics and Consent

As a bibliometric study using published documents and no human subjects, this project did not require REB approval.

Preprint

This work has been submitted as a preprint to medRxiv.

Competing Interests

Erik Driessen is Editor in Chief and Lorelei Lingard is Associate Editor for Perspectives on Medical Education. They were removed from all editorial processes relating to the handling of this submission.

Author Contributions

MA: Methodology, Formal analysis, Data Curation, Writing – Original Draft.

LM: Conceptualization, Methodology, Data Curation, Writing – Review & Editing.

HA: Methodology, Formal analysis, Writing – Review & Editing.

JC: Conceptualization, Methodology, Formal analysis, Writing – Review & Editing.

ED: Conceptualization, Methodology, Writing – Review & Editing.

KO: Methodology, Formal Analysis, Data Curation, Writing – Review & Editing, Supervision.

LL: Conceptualization, Methodology, Writing – Original Draft, Supervision.