Publishers’ and journals’ instructions to authors on use of generative artificial intelligence in academic and scientific publishing: bibliometric analysis

Publishers’ guidelines

Out of the top 100 largest academic publishers, only 24% reported guidelines for the use of GAI in research, and most of these were in the top quarter of publishers by journal count. Our analysis showed that the presence of GAI guidelines was independent of the type of publisher (table 1). Of the publishers that did have GAI guidelines, standardization was limited. Although most of the publishers cited their adherence to COPE guidelines, less than half provided links to the COPE position statement on use of AI.18 Of the publishers that did provide a link, individual guidance did not always align with the COPE statement, creating potential confusion for authors.

Despite substantial heterogeneity in publishers’ guidance, two major themes were identified. Firstly, publishers consistently prohibited GAI from being an author—namely because GAI tools cannot take responsibility for content created; a standard principle of authorship and one consistent with COPE’s position.18 Secondly, publishers encouraged the disclosure of GAI use. The nature of this disclosure varied substantially across publishers’ guidelines, such as the appropriate place for it to be cited in the manuscript. Most publishers with GAI guidelines specified which details to include in the disclosure by requesting a variety of reporting criteria, such as the model’s name, version, source, description, and usage. Elsevier provided a standardized template that included the name of the GAI tool or service used and the reason for its use.

Furthermore, the types and uses of GAI tools to which the guidelines applied varied among the publishers. For instance, while some publishers’ guidance only pertained to “AI generated text,” others also encompassed the production of images and data analysis. Several of the guidelines provided vague examples of use, such as “scholarly contributions,” “content creation,” and “preparation of a manuscript,” introducing another element of confusion for authors. Additionally, disclosure criteria for spelling and grammar raise questions around whether tools that integrate GAI technology primarily for that purpose must be held to the same standards for disclosure. This question will need to be answered, as programs such as Grammarly31 and Microsoft Office32 are implementing GAI into their spelling and grammar tools.

Finally, not all of the publishers explicitly required authors to take accountability for the output produced by GAI. This may cause confusion about the responsibility and ownership of content generated by AI tools and should be accounted for by publishers.

During the five months between our two searches of publishers’ guidelines, the number of publishers that reported guidelines increased by 40%, showing the continued interest in development of guidelines. Furthermore, the newer guidelines showed an increased emphasis on image generation and specific disclosure criteria. Although Elsevier updated its guidelines to include a section on image generation, its top journals have yet to update their guidelines to include this information even while providing a link to the publisher; a further illustration of the need for standardized guidelines to ensure proper adherence and minimal confusion for authors.

Journals’ guidelines

Most of the 100 highest ranked journals provided guidelines on the use of GAI in scientific research. Many of these journals shared the same publishers and were produced by large publishing houses, which also have guidelines and policies for GAI. Similar themes to the publishers’ guidelines applied to the guidance for journals, with great variability across guidelines and little standardization. Compared with the publishers, a lower percentage (10%) of journals linked to the COPE position statement on AI.18 Of the journals that did have a link, journal specific guidance did not entirely align with the COPE position statement. Similar to the publishers’ guidelines, the two most consistent themes were that GAI should not be listed as an author and that disclosure of the use of GAI is required. Similarly, variability was found in where to disclose this information, what details to include, and in which format. Journals published by Elsevier provided consistent guidelines requiring the use of a standard disclosure template. Additionally, across the journals, guidelines on the uses of GAI were discordant, with some specifying a combination of the writing process, image generation, or data analysis and collection, and others specifying none. Similar to the publishers’ guidelines, several journals utilized more generalized terms to describe which components of submissions were bound by the GAI guidelines. Lastly, roughly half of the journals specified that authors were accountable for the output produced by GAI. Aside from the Elsevier journals, which implemented a uniform template for disclosure, other examples of journals with well crafted guidelines included BMJ, Physiological Reviews, and PLOS ONE. These journals detailed the circumstances for which disclosure of AI is required, provided thorough and specific criteria for disclosure, including where to cite the information, and emphasized that authors must take accountability for the work resulting from use of GAI tools.

During the five months between the two searches, the number of journals reporting guidelines increased by nearly 25%, primarily those in the lower half of the top 100 highest ranked journals. Similar to publishers’ guidelines, the more recent journal guidance discussed image generation and the role of GAI in the review process. A substantially smaller percentage of the medical journals restricted their guidelines exclusively to the writing process compared with the broader list of scientific journals.

Sources of heterogeneity

We identified sources of heterogeneity in GAI guidelines among the publishers and journals, including in the dissemination of GAI related guidelines. Although some journals not only presented their own GAI guidelines but also provided a direct link to the identical publishers’ guidelines, there were instances where journals issued guidance without providing such a link. Conversely, certain journals solely provided a link to their publisher’s guidelines. This discrepancy results in a lack of centralization of information on the use of GAI. Consequently, the responsibility falls onto authors to seek out and understand the available guidelines. This setup potentially allows authors to inadvertently misuse GAI tools in scientific writing owing to an incomplete understanding of the regulations imposed by journals or publishers.

In addition to a non-centralized location for information on GAI use, we also found several instances of competing recommendations and guidance. The guidelines of some journals, such as the Journal of the American College of Cardiology, contradicted what was cited by their publisher. These inconsistencies pose challenges as authors seek out the appropriate guidelines and must decide which standards to follow.

We also found heterogeneity in the types of words used. Guidelines frequently used terms such as “disclose,” “report,” “describe,” “acknowledge,” and “document” interchangeably when instructing authors on how to present the use of generative AI in their manuscripts. This can lead to confusion, as these words have discrete definitions—for example, a disclosure of a conflict of interest is not the same as an acknowledgement of a contributor in the context of scientific publishing.

Several journals and publishers did not stipulate that authors were accountable for GAI outputs. The COPE position statement on AI asserts that authors are “fully responsible” for their work, including any portion produced by AI. This is important because, as publishers such as Elsevier and SAGE have noted, GAI can produce inaccurate, biased, or misleading outputs.1113 GAI tools are known to “hallucinate” and fabricate unfounded information.1113 Additionally, utilizing GAI tools introduces the risk of plagiarism, when text is duplicated from data sources.1233 Another element of complexity, and one acknowledged in the COPE position statement, is that AI tools are “non-legal entities.” AI tools cannot participate in matters of conflicts of interest, copyright, and license agreements. Therefore, they should not qualify as authors and take responsibility for submitted work. In fact, the issue of copyright and ownership of outputs generated by large language models will likely raise many questions that require discussion and resolution among the relevant stakeholders. Large language models are trained on vast amounts of data, potentially with various regulations and restrictions on access and sourcing. Given these concerns and the rapid adoption of GAI, publishers and journals have responded quickly to develop guidelines on proper use. Journals and publishers have recommended disclosing GAI use in the acknowledgments section. However, since AI is not human, lacks agency, and is unaccountable, there is hesitation to mention GAI in an acknowledgments section alongside collaborators.34

Heterogeneity, including the incongruence of GAI guidelines between journals and publishers, misalignment with the COPE position statement, and unclear terminology around the disclosure of GAI use, could create confusion for authors and reviewers when incorporating GAI tools into their research. A lack of clear and standardized recommendations along with frequent updates to guidelines places responsibility on authors to seek out “correct” guidance, while also diminishing the ethos of the guidance by hindering the ability of authors to appropriately follow the guidelines. Standardized recommendations would improve transparency and accountability surrounding GAI use in academia and scientific research. Although with time GAI guidelines could evolve to be discipline specific, currently during this early adoption phase authors and editors would benefit from a set of broadly encompassing, cross discipline, inclusive guidelines. A cross discipline, global initiative CANGARU (ChatGPT, generative artificial intelligence, and natural large language models for accountable reporting and use guidelines) is ongoing and the results are awaited.22 Additionally, patients may become aware of the growing use of AI in scientific research, and standardized guidelines could increase their trust in the literature.

Limitations and future directions

This bibliometric analysis represents a snapshot at six months and 10 months after the rise in popularity of ChatGPT. Guidelines were developed rapidly and specifically in response to the use of GAI and must adapt to the introduction of other GAI tools. As a result, it is likely that GAI guidance will continue to evolve as our understanding of the technology improves and as greater emphasis is placed on creating policies for GAI use, as already seen in the frequent inclusion of GAI tools besides ChatGPT in guidelines developed more recently. As GAI models continue to develop, different guidelines may need to be implemented to regulate the scope of the technology at that time. Weaknesses of the current study include the limited number of publishers and journals included. It is possible that other publishers or journals already have a higher standard of guidance on GAI use. Some publishers lacked policies on their websites, and the shared subsidiary journal guidelines that we used as proxies may not always be the perfect solution for published guidelines. Furthermore, scholarly societies associated with most of the top 100 highest ranked journals (see table 2) also play a role in the development of guidelines for GAI use. Evaluating the full effect of these societies would have been challenging given each society’s limited representation in the top 100 highest ranked journals, and as such a follow-up study evaluating their impact is needed. Another limitation is that this study was largely qualitative and therefore prone to authors’ subjectivity. However, to minimize this limitation we used a structured system of multiple reviewers as well as supervision from coauthors at several levels.

Conclusion

Substantial heterogeneity was found in guidance on the application of GAI use in academic research and scholarly writing. To our knowledge, none of the proposed recommendations were formulated through a structured consensus based guideline development process. This scenario highlights an urgent need for the establishment of cohesive, cross disciplinary policies. Such guidance should be crafted in a structured manner, integrating the perspectives of all stakeholders. This approach is crucial to counteract the Babel Tower phenomenon—that is, the confusion and lack of standardization that results from individual parties creating their own unique regulations.