Defining authorship boundaries and disclosure standards for artificial intelligence: a four-level framework for biomedical publications

Introduction The public release of ChatGPT in November 2022 introduced a novel tool to the scientific community, one that had the ability to produce academically sound and coherent text on a scale never seen before outside of human authorship. It was only weeks later that peer-reviewed papers began surfacing with ChatGPT included among the named authors. According to a report by Stokel-Walker for Nature, there were at least four other peer-reviewed papers published in early 2023 that listed ChatGPT as an author, sparking reactions from editors around the world1. This particular incident was not just a simple oversight but a fundamental problem with the concept of authorship. The ICMJE Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals specify authorship on the basis of four cumulative criteria: (1) substantial contributions to conception, design, data collection, or analysis; (2) participation in writing the paper or revising it for significant intellectual content; (3) final approval of the manuscript prior to submission for publication; and (4) willingness to take responsibility for all aspects of the work2. Authorship is thus understood as a bundle of credit and responsibility, which are logically inseparable and fundamentally irreconcilable with the nature of AI. An AI system does not bear responsibility. It cannot receive sanctions, be retracted from the literature under its name, or be judged in terms of research integrity. The discrepancy between credit and blame is an inherent and fundamental issue in AI’s contribution, as Porsdam Mann et al noted3. This paper discusses, based on the explicit statements made by prestigious journals and bioethics bodies, that the exclusion of AI from the category of formal authors is both ethically justified and realistically feasible. The paper goes further to suggest that, while it is easy to ban AI authorship in the way that most indexed biomedical journals have done, the real problem is the lack of a structured process for acknowledging the involvement of AI where this involvement is significant enough to qualify for some form of authorship credit without actually meeting that requirement. Methods This is a correspondence and policy proposal regarding authorship and disclosure of AI in biomedical research. We chose to include guidance documents and editorials published between 2023 and early 2026 from the International Committee of Medical Journal Editors (ICMJE), the Committee on Publication Ethics (COPE), and high-impact general medical journals (Nature, Science, JAMA, The Lancet, the New England Journal of Medicine, and others) that establish international standards for scholarly publishing. We summarized key statements on whether AI systems can be listed as authors, the rationale provided with reference to authorship criteria (particularly accountability and final approval), and any explicit guidance on how AI use should be reported in manuscripts. These data were organized into a descriptive table of AI authorship policies and a comparative table of AI authorship policies. Using this synthesis and building on established ICMJE and COPE principles, we developed a conceptual framework of four levels that relate the level of AI involvement in research and writing to proportionate disclosure requirements, while maintaining the absolute prohibition on AI authorship. This framework was refined iteratively to ensure internal consistency across levels and alignment with existing authorship criteria and research integrity norms. Why AI cannot satisfy ICMJE authorship criteria An analysis of the extent to which each of the ICMJE criteria aligns with the capacities of existing AI models shows a recurring trend of limited or non-existent satisfaction of criteria. While the first criterion, substantial intellectual contribution to the conception and design, is partly satisfied by current large language models (LLMs) in that they can recognize patterns within existing literature and synthesize information, these models do not create research questions via reasoning about knowledge gaps. According to Flanagin et al, writing in the journal JAMA, the former is referred to as “statistical next-token prediction,” while the latter is termed “intellectual reasoning4. Criterion II (drafting or critically revising for intellectual content) denotes the area where AI’s potential lies and, hence, is the most misused ground for claiming authorship. LLMs are very proficient in drafting grammatically correct and contextually apt scientific texts on all aspects of biomedical science5. However, the phrase mentioned by ICMJE itself – “critically revising for important intellectual content” – suggests the evaluation of ideas in terms of their truthfulness, novelty, and significance, an ability that remains beyond the scope of artificial intelligence. As Salvagno et al have warned, texts generated by AI in the realm of biomedical sciences need to be thoroughly verified by humans because LLMs may generate text that sounds convincing but is not factual at all6. Criteria III and IV are insurmountable obstacles to an AI being an author of scientific work. First, there can never be a “final approval” from an AI, for it lacks persistence in identity; it does not have an established existence recognized by the law and by organizations; it cannot provide final approval of a work because of its affiliation with such; and it cannot determine whether the manuscript, as finalized, reflects the actual findings of the research. Criterion IV – which is arguably the most important criterion – pertains to accountability2. When there is a suspicion that a paper contains fabrication and plagiarism, those involved will be held accountable. The authors bear the responsibility for the papers they produce, which includes the possibility of retraction and institutional sanctions. The lack of accountability on the part of an AI represents an existential danger to science itself, according to van Dis et al7. Converging journal policies: consensus without standardization However, regarding the authorship of articles generated by artificial intelligence, the biomedical publication industry has already taken firm action. Table 1 below provides a summary of the AI authorship policies from leading journals and indexing authorities as of 2023–2024. The editors of Nature journal made a clear statement in January 2023: “no AI tool will be accepted as a credited author on a research paper”1. The editor-in-chief of Science journal, Holden Thorp, issued an editorial stating that introducing AI-generated text into an article without disclosing it is equivalent to committing misconduct, similar to using a stolen image8. According to the updated recommendations of the ICMJE, AI tools fail to satisfy the authorship criteria, and human authors must have complete accountability for using AI-generated content2. Table 1 - Published AI authorship policies of major biomedical journals and research ethics bodies (2023). Journal/body Date of policy Position on AI authorship Disclosure requirement Reference Nature/Springer Nature 2023 AI cannot be listed as author Authors must declare AI use in Methods section Nature 2023;613:612.1 Science January 2023 AI cannot be author; no “ghostwriting” by AI Text produced by AI must be disclosed Thorp 2023.8 JAMA Network January 2023 AI/LLMs cannot be listed as authors Describe AI use in Methods; authors vouch for accuracy Flanagin et al. 2023.4 The Lancet Group February 2023 AI cannot be listed as an author (does not meet authorship criteria due to lack of accountability) AI may be used only for language/readability assistance, and any use must be transparently disclosed (typically in acknowledgements); human authors retain full responsibility The Lancet 2023.9 NEJM 2023 AI tools cannot be authors Human authors responsible for all content NEJM Editorial 2023.10 ICMJE 2023 update AI does not meet authorship criteria Humans must be accountable for AI contributions ICMJE 2023.2 COPE February 2023 AI cannot be author; cannot transfer copyright Journals must have transparent AI policies COPE 2023.11 ICMJE, International Committee of Medical Journal Editors; COPE, Committee on Publication Ethics; LLM, large language model; AI, artificial intelligence. Policies reflect published positions as of mid-2024; authors should verify current journal-specific guidance prior to submission. COPE guidelines, used by thousands of journals around the world, released an official position statement in February 2023 stating that AI cannot act as an author, cannot sign the copyright transfer or license agreement, and cannot be responsible for the accuracy or completeness of the publication11. Furthermore, COPE stressed the necessity for journals to establish and communicate clear and understandable AI policies and avoid penalizing authors who employ AI applications, provided that the latter’s usage is declared transparently. Despite the remarkable alignment of positions and recommendations, a universally accepted terminology system for specifying AI contribution levels or standardized disclosure templates has yet to be formulated. Currently, each journal publishes its unique wording for declaring AI employment, either in a dedicated Methods section, in the Acknowledgements section, or in the cover letter. This practice leads to legal uncertainties for those submitting articles to multiple journals, allows the possibility for strategic under-declaration, and creates a body of literature that varies within itself concerning the degree of contribution made by AI to the article12. As van Dis et al suggest, the lack of standardization is one of the crucial issues in the field of generative AI7. A graduated disclosure framework for AI contributions While the ban on AI’s authorship is necessary, it alone cannot be a complete response. In a scholarly setting where AI technology is now part and parcel of searching for literature (Elicit, Research Rabbit), doing statistics (Julius AI), interpreting images (GPT-4 Vision), and drafting manuscripts (ChatGPT, Claude, and Gemini), a two-fold approach to determining whether AI qualifies as an author is too simplistic3,5. According to Kitamura, such black-and-white reporting that “AI was either used or not” puts grammar-checking software and LLMs on equal footing13. In this regard, we outline a system of graduated levels of disclosure regarding the use of AI (Table 2 and the flowchart shown in Figure 1), based on the assumption that the level of disclosure should be proportional to the impact of AI on the contents of the manuscript. Level 1 (“Minimal” level) includes any use of AI in the preparation of a paper that involves tools such as spelling checks, citation management software, etc. Disclosure at this stage involves including only information about the use of AI in a short Methods section. Level 2 (“Standard” level) includes AI-assisted paraphrasing and literature review. Disclosure of AI use here should include specifying the tools used. Figure 1.: Proposed decision flowchart for determining the required level of AI disclosure in biomedical publications. Decision points progress from minimal disclosure for incidental AI use (level 1) through critical review requirements for AI-assisted data generation (level 4). The absolute exclusion of AI from authorship, independent of contribution level, is encoded as a non-negotiable terminal node, consistent with ICMJE and COPE guidance. Table 2 - Proposed graduated disclosure framework for AI contributions in biomedical publications. Disclosure level AI use category Required disclosure action Accountability burden Example Level 1 (minimal) Spell-check, grammar correction, and reference formatting Acknowledge AI tool by name in Methods or Acknowledgements Low Grammarly and EndNote Level 2 (standard) Text paraphrasing, literature summarization, and basic data visualization Name AI tool, version, date of use, and general purpose in Methods Moderate ChatGPT and Gemini for literature review Level 3 (enhanced) Generation of draft manuscript sections, image creation, and code writing Detailed Methods subsection: tool, version, prompt strategy, human review process High GPT-4 for initial draft of Discussion Level 4 (critical) Data generation or transformation, statistical analysis, and diagnostic interpretation Full supplementary disclosure; independent human verification of all AI outputs mandatory Very High AI-assisted histopathology diagnosis Never Conception of research question, critical intellectual analysis, final manuscript approval, and ethical accountability AI cannot be listed as author under any circumstance Absolute Prohibited per ICMJE, COPE, and all major journals Framework informed by ICMJE 2023 Recommendations, COPE Position Statement on AI (2023), and published journal policies. “Never” row reflects the universal consensus that AI cannot be listed as an author regardless of contribution level. Importantly, what the framework makes explicit is something that no amount of disclosure could ever justify: AI taking on the functions specified in ICMJE criteria III and IV. Disclosure cannot turn AI use into authorship, since the problem is not a lack of transparency but rather the ontological problem of a lack of the moral accountability required for authorship. This problem is formulated by Hosseini et al as a question of “moral agency”: authorship assumes a subject who can assume moral responsibility for her/his claims, and AI lacks this form of moral agency14. Ethical implications: integrity, reproducibility, and equity Apart from the issue of authorship itself, there are multiple interrelated issues with incorporating AI into biomedical research that have implications for research integrity, reproducibility, and equity. First, there is the problem of “hallucination,” whereby LLMs can produce syntactically plausible yet factually incorrect outputs, such as made-up citations9. Sallam et al found that the output of LLMs can reproduce and amplify the existing biases in the training dataset, suggesting that research conducted with AI assistance might exacerbate inequalities in disease representation among demographic groups15. Second, reproducibility, a hallmark of biomedical sciences, will not be achieved when the contribution of AI is not transparently reported. When a researcher employs the help of an AI system at any step of analysis, whether it is code generation or interpreting the output, and does not mention the version of the tool, prompts used, or temperatures, replication will not be possible. There is a further complication with LLMs, which are non-deterministic: submitting the same prompt twice to the same model will not produce the same output3,16. For publications aiming at openness and data sharing, this implies a need for a novel definition of “methods sufficiency.” It is also important to consider the equity considerations of AI-enabled writing. Access to high-capacity LLMs, especially premium-level LLMs, is not uniformly available across organizations and jurisdictions. Scholars at well-endowed organizations in wealthy nations can leverage AI to rapidly generate articles, prepare peer reviews, and compose funding applications, whereas scholars at poorly endowed organizations cannot gain access to similar technology17. Lund and Wang contended that such an imbalance would exacerbate disparities in academic productivity and publications, thereby biasing the geographical and demographic composition of biomedical knowledge5. Recommendations We make the following evidence-based recommendations for journals, institutions, funders, and authors: For journals: Adopt the graduated disclosure framework proposed in Table 2, incorporating a structured AI Disclosure Statement as a mandatory submission field, distinct from the Acknowledgements section. Publish a clear, accessible AI policy and update it annually as the technological landscape evolves. For institutions: Develop institutional AI use policies for research that define permissible and impermissible AI applications, provide guidance on training data provenance and data privacy, and establish clear procedures for handling suspected AI misconduct in grant applications and publications. For funders: Require AI disclosure statements in grant progress reports and final outputs. Consider mandating that AI tools used in funded research be specified by name, version, and provider to enable independent auditing. For authors: Document AI tool use contemporaneously throughout the research process, not retrospectively at the point of submission. Maintain a clear record of which manuscript sections were generated, revised, or checked with AI assistance, and submit this record upon editorial request. For the ICMJE: Develop and publish a companion guidance document specifically addressing AI in the context of each of the four authorship criteria, providing illustrative examples of qualifying and non-qualifying AI use to reduce interpretive ambiguity across journal editorial teams. Conclusion Generative AI represents a significant paradigm change in biomedical science research. There is no dispute that AI should not be considered an author, but now the problem arises in developing a uniform system of appropriate disclosure across a range of applications. This step-by-step proposal allows one to introduce AI into science in an ethical way without restricting the possibilities offered by such technology. It is especially important for the field of biomedical science due to its direct connection to clinical practice and policymaking.