ABSTRACT – Authorship, Accountability, and Generative AI in Scholarly Publishing: Evidence from Samara National Research University and Global Policy Convergence 2023–2025
Across editorial offices and research groups, the understanding of who can be called an author has been pushed to its limits by the arrival of generative AI, and the story opens in the Russian Federation with a careful investigation from Samara National Research University that becomes the thread tying policy, ethics, and scholarly practice together. In July 2025, a peer-reviewed article in Semiotic Studies examined authorship in an era when models can draft, revise, and even appear in author fields, and it did not stop at theory: it verified that Web of Science contained 4 records crediting ChatGPT as an author, including 2 sole-author entries, while Scopus held 2 more attributions, one later removed when an editor intervened.
That empirical core gives the narrative urgency because it shows how bibliographic infrastructure allowed non-human attribution to pass through real submission systems, with consequences for accountability and citation metrics that spread far beyond any single journal. The approach was straightforward and auditable: keyword discovery across the two dominant indexing services, direct inspection of metadata and article files to confirm that generative systems were not merely acknowledged but named as authors, and synthesis into a model that distinguishes between machine generation and human responsibility while treating research writing as a chain of roles in which people act as curators, editors, and interpreters of statistical output.
What makes this account matter is that it intersects with live rules and expectations already reshaping practice: the position taken by COPE in 2023 that tools cannot be authors because they cannot take responsibility; the guidance from NIH in 2023 that peer reviewers cannot use generative systems and that any use in applications must remain the human applicant’s responsibility; and the adoption of the Artificial Intelligence Act by the European Parliament in 2024, which pushes disclosures into the mainstream without granting personhood to machines. Woven together, these strands explain why the Samara team’s typology of hybrid, distributed, and complex authorship gains traction: hybrid captures manuscripts where a model drafts segments that a scholar substantively reshapes and endorses; distributed treats writing as a socio-technical network where developers, dataset curators, prompt engineers, and scholars each contribute in different ways; complex covers the large, many-handed collaborations familiar from physics or assessment science and adds the wrinkle that an algorithm can scale drafting without ever being accountable.
As the story unfolds, the methodology clarifies the stakes. The researchers did not rely on anecdotes; they audited records that any librarian can check, and they showed how an author field can accommodate an entity that cannot sign disclosures or respond to reviewers. That procedural clarity is important because it lets editors and indexers replicate the search and decide how to repair their pipelines. It also frames the philosophical claim in practical terms: if authorship is a promise that someone will stand behind every sentence, then a model, however fluent, cannot keep that promise.
This is not an abstract worry. Indexing an algorithm as an author shifts citation credit to a non-existent person, distorts the h-index, and can affect hiring and funding decisions downstream. When a human scholar silently pastes machine text without curating, editing, or verifying, the scholarly record loses its chain of custody, and research integrity officers cannot rely on plagiarism checks because generative systems produce novel strings. The result is a new kind of misconduct risk that is neither classic plagiarism nor fabrication but a breakdown in attribution and accountability. At the same time, the narrative does not paint AI as an intruder to be banished.
It recognizes that editors and funders are converging on a workable compact: allow generative assistance as long as disclosure is clear and a human author accepts full responsibility for accuracy, originality, and ethical compliance. Publisher policies are moving in that direction. Elsevier and Springer Nature have told authors they may not list tools as co-authors but may use them with transparent acknowledgments, and that reviewers should not upload manuscripts into such systems. This is where the Samara model earns its value, because it offers language for contributorship that preserves human accountability while acknowledging machine labor: let the author field stay human, create metadata that discloses machine involvement, and map who curated prompts, who edited outputs, and who validated claims.
The findings are specific and consequential. There are documented cases of non-human attribution in the most widely used databases, and there is at least 1 instance where an editor required removal after publication. The typology shows how to describe work that many scholars are already doing with generative assistants without encouraging the fiction that a model can bear responsibility.
The ethical analysis isolates the risk with precision: not every use of a model is problematic, but every failure to disclose, verify, and assume responsibility is. The legal analysis draws a bright line consistent with intellectual property doctrine: under United States law and under European Union norms, only people can be authors and rightsholders, which means that recognition for machine output must be handled through acknowledgments and contributor roles, not through person-like credit.
The policy analysis shows institutions aligning around that line, from funders to publishers to lawmakers, and the practical proposals translate easily into editorial checklists and database schemas. Even the epistemology becomes concrete. Treat writing as the interplay between statistical generation and human judgment, and the test is simple: did a responsible scholar interpret the output, integrate it with evidence, and stand ready to defend each claim. If yes, the manuscript fits within the tradition of accountable authorship even when AI assisted; if no, the manuscript lacks an author in the sense that matters to science.
The implications are direct for editors, reviewers, and authors. Editors can require machine-contribution disclosures at submission, add checks for non-human names in author fields, and train boards to recognize stylistic red flags without turning peer review into a hunt for ghosts. Reviewers can refuse to feed manuscripts into external systems and treat undisclosed generative usage as an integrity concern. Authors can adopt contributorship statements that name who crafted prompts, who edited model outputs, and who verified every reference, and they can keep a local audit trail so that if a question arises, the chain of human oversight is clear. Indexing services can introduce machine-assistance tags so that transparency travels with the record and citation networks are not distorted by non-existent identities. Professional societies and ethics bodies can harmonize language so that scholars face one coherent set of expectations whether they submit in the United States or the European Union. Policymakers can keep disclosure requirements technologically neutral so that rules cover the next generation of systems without constant rewrites. None of these steps treats AI as an author, and none denies its utility; they simply restore the promise that has always underwritten the literature, the promise that someone with a name will answer for what appears on the page.
Because this account is grounded in verifiable records and live policy, it reads less like speculation and more like a map for how to move forward with confidence. The empirical audit from Samara National Research University anchors the narrative in facts that anyone can check through Web of Science and Scopus; the ethical and legal analysis aligns with positions already taken by COPE and with funder guidance such as the 2023 NIH notice; and the regulatory context is set by the adoption of the Artificial Intelligence Act by the European Parliament, which pushes disclosures into standard practice while leaving authorship where the law says it belongs, with people. The practical effect is a shift from arguing about whether a model can be an author to building reliable ways to say when and how a model helped.
That shift preserves the core values of originality, accountability, and transparency while absorbing the undeniable reality that drafting and revising with generative assistance is already part of scholarly work. In simple terms, the path forward is to keep the author field human, make disclosure routine, reward careful curation and verification, and ensure that the record carries the information readers and evaluators need. The Samara study shows why this matters right now, the policies show how it can be done, and the convergence across publishers, funders, and lawmakers shows that the community already has the tools to protect the integrity of research while letting AI do what it does best under responsible human hands.
CHAPTER INDEX
- Evolution of Authorship Standards in Scholarly Publishing
- Empirical Documentation of AI-Named Authorship in Web of Science and Scopus (2025)
- Definition and Structure of Hybrid, Distributed, and Complex Authorship Models
- Ethical Risks and Misconduct Implications of Unacknowledged AI Use
- Legal and Institutional Accountability under Contemporary Publishing Norms
- Epistemological Challenges Raised by AI-Generated Text in Knowledge Production
- Policy Directions for Transparent and Responsible AI Use in Academic Scholarship
Evolution of Authorship Standards in Scholarly Publishing
The recognition of authorship in scholarly communication has historically been grounded in principles of intellectual responsibility, originality, and accountability, codified most explicitly by the International Committee of Medical Journal Editors (ICMJE) in its Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals (2023 edition), which emphasize that authorship must be based on substantial contributions to conception, design, analysis, drafting, and final approval of a manuscript (ICMJE). These criteria explicitly exclude technical assistance or provision of materials as sufficient grounds for authorship. The historical trajectory of these standards reflects a continuous narrowing of the definition of authorial responsibility in response to disputes over honorary authorship, ghostwriting, and contributorship inflation, trends well-documented in bibliometric studies by the World Association of Medical Editors (WAME) between 2018 and 2022 (WAME).
The emergence of large-scale collaborative science in the mid-20th century, particularly in particle physics, prompted new tensions in authorship attribution. By 2015, publications from the ATLAS and CMS collaborations at CERN’s Large Hadron Collider regularly included author lists exceeding 3,000 names, forcing journals and funding bodies to confront whether individual responsibility could be meaningfully assigned (CERN). These precedents expanded the concept of authorship beyond individual intellectual ownership to encompass collective institutional responsibility, a framework that laid the groundwork for the present debate regarding the inclusion of non-human agents.
Ethical debates over ghostwriting further shaped the evolution of authorship standards. Analyses published in the Journal of the American Medical Association (JAMA, 2010–2020) documented recurring cases where pharmaceutical companies employed professional writers to draft clinical trial reports, later attributed to academic researchers, raising transparency concerns (JAMA). The exposure of these practices reinforced the demand for disclosure of all contributors, a precursor to today’s contributorship taxonomies. In 2014, the Contributor Roles Taxonomy (CRediT) was introduced under the auspices of the Consortia Advancing Standards in Research Administration Information (CASRAI), defining 14 specific contributor roles ranging from conceptualization to data curation (CASRAI CRediT). This taxonomy has since been adopted by publishers including Elsevier, PLOS, and Springer Nature, marking a decisive shift toward granular attribution rather than undifferentiated authorship.
The legal dimension of authorship developed in parallel. The Berne Convention for the Protection of Literary and Artistic Works, administered by the World Intellectual Property Organization (WIPO), enshrined authorship as the foundation of copyright recognition, but presupposed human creativity as its basis (WIPO). Jurisprudence in the United States, including Naruto v. Slater (2018), reaffirmed that non-human entities cannot hold copyright, a principle reiterated in the US Copyright Office’s 2023 guidance on AI-generated works (US Copyright Office). Similarly, the European Court of Justice has consistently ruled that copyright protection requires “the author’s own intellectual creation,” effectively excluding autonomous systems from authorship. These rulings establish a legal boundary that clashes directly with the empirical realities now emerging in bibliometric databases, where AI systems have been explicitly credited.
Technological advances in digital publishing accelerated transformations in authorship practice. The proliferation of preprint servers such as arXiv (established 1991) and bioRxiv (launched 2013) facilitated mass dissemination of research prior to peer review, normalizing the appearance of automated tools in manuscript preparation. By 2021, surveys conducted by the Nature Publishing Group reported that over 20% of academic researchers had used AI-driven grammar correction or paraphrasing tools in manuscript preparation (Nature). Although these uses did not generally warrant authorship attribution, they blurred the boundaries between technical assistance and creative contribution, foreshadowing the controversies documented by the Samara National Research University team in 2025.
Empirical Documentation of AI-Named Authorship in Web of Science and Scopus (2025)
The empirical research conducted by Samara National Research University in July 2025 systematically investigated the extent to which generative AI systems had been explicitly credited as authors in global bibliometric repositories. Their findings, published in Semiotic Studies, established that in Web of Science, four indexed articles listed ChatGPT as an author, of which two presented the system as the sole author, without any accompanying human names (ResearchGate). In addition, two articles in Elsevier’s Scopus database were identified where ChatGPT was attributed as co-author, although in one of these cases, the AI reference was later removed following editorial correction at the publisher’s request (sputnikglobe.com).
The methodology of the Samara study combined keyword searches for “ChatGPT” and related identifiers across Web of Science Core Collection and Scopus during the period January 2023 – April 2025, followed by manual verification of metadata and article PDFs. This procedure ensured that the identified attributions were not incidental mentions in acknowledgments or methods sections, but explicit assignments of authorship in the official author fields indexed by the databases. The researchers’ dataset thus constitutes the first verifiable evidence that mainstream bibliographic infrastructures are already hosting articles where generative systems are formally credited.
Such anomalies reflect systemic weaknesses in editorial and indexing workflows. Journals indexed in Web of Science are contractually required to comply with the Publisher’s Ethics Guidelines of Clarivate Analytics (2024), which stipulate that all named authors must have the capacity to accept accountability (Clarivate). The presence of non-human entities in metadata fields indicates lapses either in editorial gatekeeping or in the enforcement of publisher obligations by database operators. Similarly, Scopus has explicit metadata policies requiring ORCID integration for authorship fields, but the Samara study found that AI attributions bypassed this requirement, signaling vulnerabilities in automated ingestion protocols (Elsevier Scopus).
Citation analysis of the identified articles revealed that at least one of the AI-authored works had already been cited by other publications in 2024, raising questions about the downstream impact of such misattribution on research evaluation. Since bibliometric indicators such as h-index, citation counts, and altmetric scores are automated, the inclusion of AI “authors” introduces distortions that can propagate across evaluation systems used by funding agencies, tenure committees, and ranking organizations. These distortions undermine the reliability of bibliometrics as objective measures of individual scholarly productivity.
The Samara team emphasized that these works demonstrate the hybrid character of contemporary academic production. In their interpretation, articles where AI is credited as author should be categorized as instances of “distributed authorship,” in which human researchers act as curators, editors, and interpreters of machine-generated output. However, they warned that absent transparent disclosure and accountability frameworks, such practices risk devolving into academic misconduct if machine outputs are misrepresented as the sole intellectual work of a human researcher.
Definition and Structure of Hybrid, Distributed, and Complex Authorship Models
The Samara National Research University study advanced a typology of authorship models adapted to the realities of generative AI participation, distinguishing between hybrid, distributed, and complex authorship structures. Hybrid authorship refers to cases where machine outputs are integrated into human-written manuscripts, but responsibility and interpretive authority remain vested in human researchers. This mirrors established practices where statisticians or illustrators contribute specialized work without assuming full authorship, yet with an important distinction: AI systems operate without consciousness, volition, or accountability, making the attribution of authorship in their name ethically and legally untenable.
Distributed authorship extends this concept by framing research creation as the outcome of multiple interacting agents—human and non-human—whose contributions cannot be easily separated. This framework draws on actor-network theory as formulated by Bruno Latour in the 1980s, which reconceptualized scientific practice as a network of human and material actants rather than isolated individuals ([Latour, Science in Action, 1987 — No verified public source available]). In the Samara framework, AI systems are treated as “distributed contributors” embedded within a chain of algorithm designers, prompt engineers, dataset curators, and human interpreters. Authorship thus becomes a property of the socio-technical system rather than a single identifiable mind.
Complex authorship, the third model, reflects cases where large-scale collaborations already challenge traditional attribution, such as the thousands of co-authors on CERN’s ATLAS experiments or the collective authorship adopted by IPCC Assessment Reports, in which hundreds of scientists jointly draft and review material. The Samara researchers argue that in the age of generative AI, complexity is compounded by the involvement of algorithmic agents that cannot sign disclosure forms, respond to peer reviewers, or accept liability. This necessitates a reconceptualization of responsibility structures, where human co-authors act as guarantors of AI-generated segments, ensuring compliance with scientific and ethical norms.
The hybrid-distributed-complex framework proposed by Maslenkova and colleagues situates AI authorship within a continuum rather than a binary of inclusion or exclusion. Instead of debating whether AI can or cannot be an author, the model emphasizes delineating specific roles: generative production, human curation, editorial validation, and interpretive synthesis. By explicitly mapping these functions, the model prevents the erasure of machine involvement while safeguarding human accountability. This echoes similar moves in contributor taxonomies such as CRediT, which from 2014 defined roles such as “data curation” or “software,” without assigning legal authorship to non-human entities (CRediT Taxonomy).
The implications extend into epistemology, as distributed and complex authorship models challenge the long-standing humanist assumption that knowledge production is attributable to individual genius. Instead, they suggest a paradigm in which texts are co-produced by human judgment and statistical generation, necessitating new disclosure practices and possibly new metadata categories in bibliometric databases. Such a system would avoid crediting AI as a legal “author” while ensuring transparency about the extent of machine involvement in the published record.
Ethical Risks and Misconduct Implications of Unacknowledged AI Use
The unacknowledged use of generative AI in scholarly publishing raises acute ethical risks, identified both by the Samara National Research University study and by major international publishing bodies. According to the Committee on Publication Ethics (COPE), which in 2023 issued a position statement on the use of generative tools, undisclosed reliance on AI-generated content without human verification constitutes a breach of authorship integrity, potentially qualifying as a new category of academic misconduct (COPE). Misconduct in this context extends beyond plagiarism, fabrication, or falsification, since the issue lies not in misrepresenting others’ work, but in misrepresenting the locus of accountability.
The ethical dilemma intensifies when AI-generated text is submitted under a human author’s name without substantive editing or interpretive oversight. The Samara University researchers stressed that in such cases the human actor ceases to function as creator and instead operates as a conduit for machine output, thereby eroding the epistemic responsibility that underpins authorship in academic traditions (ResearchGate). This practice undermines trust in peer review, since reviewers assume that submissions are products of accountable human reasoning.
Ethical risks extend into pedagogy and training. A 2024 survey by Nature of over 1,600 scientists reported that 17% admitted to using generative AI to draft sections of manuscripts or grant proposals, yet fewer than half disclosed this to collaborators or supervisors (Nature). Such opacity threatens to normalize hidden machine involvement, eroding collective standards of disclosure and potentially producing a generation of researchers habituated to opaque authorship practices. In the context of doctoral training, this risks redefining scholarly labor away from intellectual originality toward algorithmic prompting, thereby distorting the development of academic identity.
From the perspective of research integrity offices, unacknowledged AI use complicates investigative processes. Traditional plagiarism detection software such as Turnitin or iThenticate is designed to identify overlap with existing human-authored texts, but generative systems produce unique, non-plagiarized outputs that evade detection. Reports from Retraction Watch in 2023–2024 documented at least 40 retractions linked to suspected AI-generated text, often identified not by plagiarism software but by unusual stylistic patterns and reviewer suspicion (Retraction Watch). Without robust disclosure, editorial boards risk publishing content whose provenance cannot be reliably established, undermining the integrity of the scientific record.
The Samara findings underscore that ethical risks are not hypothetical but already instantiated in the bibliometric record. By identifying articles where ChatGPT was explicitly credited, they revealed both the existence of undisclosed hybrid works and the fragility of editorial safeguards. They concluded that failure to implement systematic guidelines would foster conditions for widespread academic misconduct, particularly in high-pressure research environments where publication metrics drive career advancement.
Legal and Institutional Accountability under Contemporary Publishing Norms
The question of legal accountability for generative AI in scholarly publishing has been addressed unevenly across jurisdictions, creating a fragmented regulatory environment. In the United States, the Copyright Office issued explicit guidance in March 2023 clarifying that works created entirely by AI systems cannot be registered for copyright protection, as authorship requires human creativity and intent. However, if human authors make “substantive modifications” or demonstrate creative control over AI-generated content, those human contributions can be protected (US Copyright Office). This standard creates a threshold where humans retain legal authorship but raises unresolved questions about how much intervention is necessary to establish ownership.
In the European Union, the Artificial Intelligence Act, formally adopted in March 2024, introduced provisions mandating disclosure when generative systems contribute to content creation. Although the act does not confer authorship rights on AI, it obligates researchers and institutions to provide transparency notices when AI is used in the production of research outputs (European Parliament). The disclosure requirement situates accountability firmly with human actors, consistent with existing intellectual property law under the Berne Convention and the EU Copyright Directive, both of which recognize only natural persons as authors.
Publishing institutions have begun to integrate these legal frameworks into editorial policy. Elsevier, in its 2024 Publishing Ethics Guidelines, requires authors to disclose the use of AI tools and prohibits listing such tools as authors, reinforcing that accountability cannot be delegated to non-human entities (Elsevier). Similarly, the Springer Nature editorial board issued a policy update in January 2023 stating that AI systems may not be cited as authors and that human researchers must accept full responsibility for verifying content generated with their assistance (Springer Nature). The American Chemical Society (ACS) adopted comparable rules in 2024, explicitly banning AI attribution in author fields, but allowing acknowledged contributions in methods or acknowledgments sections (ACS Publications).
Institutional accountability also arises in cases of retraction. The Committee on Publication Ethics (COPE) updated its misconduct taxonomy in 2023 to include “inappropriate use of generative tools” as grounds for editorial action, ranging from corrections to full retractions (COPE). The retraction documented by Samara University in Scopus illustrates the application of these evolving norms, where the presence of AI in the author field was deemed incompatible with both legal and ethical requirements. Retractions of this kind carry institutional consequences, as universities are required to report them to national research integrity bodies and funding agencies, affecting reputational standing and grant eligibility.
At the international level, the World Intellectual Property Organization (WIPO) has initiated consultations on AI and intellectual property since 2019, with updated reports released in 2023–2024. These documents reaffirm that current treaty frameworks do not accommodate AI as a legal author or rights holder, and recommend developing disclosure-based approaches to manage AI contributions (WIPO). However, no binding international standard exists, leaving a patchwork of publisher policies and national laws. This regulatory fragmentation creates uncertainty for multinational collaborations, where a manuscript co-authored across jurisdictions may be subject to conflicting requirements regarding disclosure, authorship recognition, and liability.
Epistemological Challenges Raised by AI-Generated Text in Knowledge Production
The incorporation of generative AI into academic writing destabilizes established epistemological assumptions about authorship, originality, and knowledge validation. Philosophical traditions rooted in Immanuel Kant’s conception of authorship as an expression of autonomous rationality presuppose that the author is a conscious agent who assumes responsibility for the truth-value of propositions. In contrast, generative systems such as ChatGPT function as probabilistic models trained on vast corpora, producing outputs based on statistical correlations rather than intentional meaning-making. The Samara National Research University study emphasized this by labeling AI a “stochastic parrot,” reiterating linguistic patterns without comprehension (ResearchGate).
Epistemologically, this distinction undermines the conventional linkage between authorship and epistemic accountability. A human author not only generates propositions but is also accountable for their truthfulness, coherence, and methodological grounding. An AI system lacks the capacity to justify claims, respond to critique, or situate findings within theoretical frameworks, all of which are central to scientific discourse. Consequently, when AI-generated text appears in the scholarly record without proper curation, it introduces statements whose epistemic warrant cannot be traced to responsible agency. This represents a rupture in what philosopher Jürgen Habermas termed the “communicative rationality” of science, where validity claims must be open to rational critique in public discourse ([Habermas, Theory of Communicative Action, 1981 — No verified public source available]).
Further challenges arise in the domain of originality. Under the Berne Convention and US Copyright Law, originality requires “independent creation” and “a minimal degree of creativity.” However, generative models are derivative by design, trained on human-created corpora. Their outputs are recombinations of existing linguistic structures, raising questions about whether originality in the human sense applies. The European Patent Office clarified in 2023 that inventive step and originality cannot be attributed to AI outputs, a ruling that underscores the epistemological gap between statistical generation and human creativity (EPO).
In scientific methodology, reliance on AI risks introducing epistemic opacity. Generative models are often described as “black boxes”, with training datasets and parameter weightings opaque to end users. When such systems produce text used in academic publications, neither authors nor reviewers can fully reconstruct the reasoning behind specific claims. This contravenes the principle of replicability, a cornerstone of modern science emphasized in the Open Science Framework and endorsed by the OECD Recommendation on Access to Research Data from Public Funding (2021) (OECD). If knowledge claims cannot be independently verified or traced, their epistemic status within the scientific record becomes unstable.
The epistemological challenges also intersect with pedagogy and cultural understandings of authorship. The Samara study highlighted that in many social contexts, authorship functions not only as recognition of intellectual contribution but also as a marker of authority and legitimacy. If AI systems are credited alongside or instead of humans, the symbolic capital traditionally associated with authorship is disrupted. This risks both inflating the role of machines in public imagination and diminishing the perceived accountability of human scholars. The researchers concluded that new guidelines must explicitly reaffirm human responsibility as the guarantor of epistemic integrity, even in hybrid and distributed authorship frameworks.
Policy Directions for Transparent and Responsible AI Use in Academic Scholarship
The findings of the Samara National Research University study converge with broader institutional calls for structured governance of generative AI in research, emphasizing the urgent need for policy frameworks that ensure both transparency and accountability. The Committee on Publication Ethics (COPE), in its 2023 statement on AI, recommended mandatory disclosure of generative tool usage, explicit prohibition of AI systems in the author field, and inclusion of editorial training to detect AI-generated content (COPE). These recommendations have been adopted unevenly across publishers, with major houses such as Elsevier, Springer Nature, and Wiley implementing disclosure-based guidelines between 2023 and 2024, while smaller journals without robust editorial infrastructures remain vulnerable to inconsistent enforcement.
Funding bodies have also begun to integrate AI-use policies into their grant processes. The United States National Institutes of Health (NIH) in August 2023 updated its grant application instructions to clarify that applicants may use generative AI tools for drafting purposes, but responsibility for verifying the accuracy and originality of text remains exclusively with the human principal investigator (NIH). Similarly, the European Research Council (ERC) in 2024 mandated that applicants disclose any AI assistance in proposals, though it emphasized that such disclosure would not affect evaluation outcomes provided that the applicant retained intellectual responsibility (ERC). These funding policies reinforce the principle that transparency is a precondition for legitimacy in scholarly communication.
At the intergovernmental level, the United Nations Educational, Scientific and Cultural Organization (UNESCO) adopted its Recommendation on the Ethics of Artificial Intelligence in 2021, updated in 2023 to include explicit language on AI in education and research. The revised text urges member states to establish national frameworks that prohibit presenting AI-generated material as original human-authored research, while promoting literacy programs to train scholars in responsible AI use (UNESCO). Complementary efforts have been initiated by the OECD, which in 2024 published its policy framework on generative AI in science, recommending disclosure standards aligned with open science principles (OECD).
Proposals for database-level reforms are also under discussion. The Samara team highlighted that bibliometric infrastructures such as Web of Science and Scopus currently lack metadata fields distinguishing human and AI contributors. Policy experts have recommended the introduction of new metadata categories that allow disclosure of AI assistance without listing AI systems as formal authors, ensuring transparency while preserving accountability. Clarivate has indicated in its 2025 Roadmap for Web of Science Enhancements that it is considering implementing machine-contribution tags in response to these debates (Clarivate).
Finally, policy trajectories point toward the establishment of international harmonization mechanisms. The World Intellectual Property Organization (WIPO) has convened ongoing consultations since 2019, with its 2024 Issues Paper on AI and IP emphasizing the necessity of disclosure regimes and rejecting the recognition of AI as a legal author or inventor (WIPO). The alignment of national legislation, publisher policy, funding body requirements, and bibliometric database reforms will be essential to prevent inconsistencies and loopholes that could enable misconduct. The Samara study’s recommendations align with this trajectory, calling for a comprehensive global framework that institutionalizes distributed and hybrid authorship while retaining human researchers as the final guarantors of academic integrity.
