A consensus has crystallised across the scholarly publishing and research-funding landscape over the past two years: generative artificial intelligence tools cannot be listed as authors on a research output, but their use must be disclosed. What remains unsettled — and what is now the focus of active AI authorship disclosure research among journals, universities, and funders — is exactly how that disclosure should be structured, who is responsible for verifying it, and how institutions distinguish genuine human intellectual contribution from AI-assisted production of text, code, data analysis, or images.
The policy convergence is real. ICMJE guidance, COPE position statements, and publisher-level policies from major scholarly houses all now hold that large language models and generative AI systems fail the basic test of authorship: they cannot take responsibility for a work’s accuracy, cannot agree to be accountable for its integrity, and cannot hold the legal or ethical liability that authorship implies. What has proliferated instead is a patchwork of disclosure mechanisms — acknowledgement sections, dedicated “AI use” statements, methods-section declarations, and in some cases structured metadata — with no single format yet dominant. That divergence is precisely where contributor role frameworks are proving useful.
Why AI Authorship Disclosure Research Is Accelerating in 2026
Three pressures are driving institutions to formalise their approach faster than in previous years. First, the EU AI Act’s phased compliance timeline is pushing research-performing organisations to document AI use in a way that satisfies both scholarly integrity norms and emerging regulatory transparency obligations. Second, UK institutions preparing for the REF 2029 cycle are under pressure to demonstrate that submitted outputs meet originality and integrity standards that predate generative AI, which means research offices need defensible, auditable disclosure practices now rather than in 2028. Third, funders are beginning to ask more precise questions in grant reporting about how AI tools were used in proposal writing, data analysis, and manuscript preparation — a shift that reflects broader AI regulation research funding bodies are grappling with as they update terms and conditions.
The practical effect is that “disclose AI use” is no longer sufficient as a policy statement. Research offices, journals, and funders are being asked to specify: disclose what, at what level of granularity, verified by whom, and recorded where. This is the gap that structured contributor taxonomies were originally built to close for human contributions — and it is why they are now being extended, cautiously, into AI governance conversations.
Contributor Role Frameworks as a Principled Dividing Line
The core conceptual tool available to institutions is not new. CASRAI originated the CRediT contributor role taxonomy in 2014. The standard is now stewarded by NISO as ANSI/NISO Z39.104-2022, and it defines fourteen discrete contribution types — from Conceptualization and Methodology to Writing – original draft and Writing – review and editing — each of which can be attributed to a named, accountable human contributor.
The taxonomy’s original purpose was to disaggregate authorship credit among multiple humans on a single paper. Its underlying logic, however, generalises usefully to the AI question: authorship requires accountability for a specific, nameable contribution, and accountability requires an agent capable of bearing responsibility. A generative AI tool can plausibly be described as having assisted with tasks that map onto CRediT categories such as drafting text, generating code, or performing data curation — but it cannot occupy the CRediT role itself, because a role assignment implies the assignee can answer for the work’s validity under scrutiny, including retraction proceedings, correction requests, or research-integrity investigations. This is the principled basis journals and institutions are increasingly citing: the same contributor-role logic that separates a data-generating instrument from the humans who interpreted its output can separate an AI writing or coding assistant from the humans who directed, checked, and take responsibility for its use.
Several publishers now ask authors to describe AI involvement in language that echoes CRediT categories — for example, specifying that a tool was used to support Writing – original draft but that Conceptualization, Formal analysis, and Writing – review and editing remained entirely human. This is a productive middle path: it does not require a new taxonomy, but it borrows the existing one’s granularity to make disclosure statements auditable rather than vague.
Where Implementation Diverges
Despite the shared principle, practical implementation varies considerably across the research ecosystem:
- Placement of disclosure. Some journals require AI use statements in the methods section (treating it as a methodological detail); others require a separate acknowledgements-adjacent declaration; a smaller number embed it in cover letters reviewed only by editors, not readers.
- Granularity required. Some policies accept a blanket statement (“generative AI was used to improve language and readability”); others, more aligned with CRediT-style precision, require task-level specification of which sections or functions involved AI assistance.
- Tooling identification. A minority of policies require naming the specific tool and version used, which matters for reproducibility and for tracking model-specific error patterns, but raises practical questions when authors use multiple tools across a long project.
- Verification mechanisms. Almost no institution has a reliable technical means of verifying that disclosed AI use is complete and accurate; disclosure remains largely an honour system underpinned by researcher attestation, similar to conflict-of-interest declarations.
- Funder versus publisher scope. Funders such as UKRI and participants in cOAlition S are beginning to address AI use in grant terms, but their focus tends to sit upstream — on proposal preparation and data management plans — whereas publisher policies focus downstream, on the submitted manuscript. Institutions sitting between the two face a compliance gap where neither policy layer fully covers the research lifecycle.
This divergence is not simply inconsistency for its own sake; it reflects genuinely different institutional risk profiles. A journal’s primary concern is the integrity of the published record. A funder’s primary concern is the integrity of the proposal and reporting process. A university’s research integrity office must satisfy both, plus internal disciplinary and REF-adjacent obligations, which is why many research offices are now building disclosure requirements that exceed the minimum asked by any single external stakeholder.
What This Means for Research Administrators
For research administrators, the practical task is less about resolving the philosophical authorship question — that consensus is largely settled — and more about operationalising disclosure consistently across a diverse portfolio of journals, funders, and disciplines. Several concrete steps follow from the analysis above:
- Adopt CRediT-aligned language in institutional AI-use disclosure templates, so that researchers describe AI assistance using the same task-level vocabulary already familiar from authorship contribution statements, rather than inventing a parallel, less precise disclosure format.
- Build AI disclosure into existing research integrity and authorship training rather than treating it as a standalone policy, since the underlying skill — accurately attributing who or what did what — is the same competency CRediT training already builds.
- Track the funder-versus-publisher compliance gap explicitly in grant management workflows, particularly where UKRI or Horizon Europe-funded projects will also be submitted to journals with independent AI disclosure requirements.
- Maintain records of AI-use disclosures in a form that could support a future research-integrity enquiry, given that verification remains attestation-based and institutions, not authors alone, may be asked to demonstrate due diligence.
- Monitor evolving guidance from ORCID, DataCite, and CrossRef on whether AI-tool disclosure will eventually be captured as structured, machine-readable metadata rather than free-text statements — a shift that would materially change how research offices audit compliance at scale.
This agenda sits squarely within the broader landscape of generative AI policy research institutions are now expected to maintain, alongside data management, open access, and research integrity policy suites. It also intersects with wider questions of AI ethics academic institutions face around equitable access to AI tools, disclosure burden on early-career researchers, and the risk that inconsistent policy enforcement disadvantages authors publishing across journals with conflicting requirements.
A Settled Principle, an Unsettled Practice
The authorship question itself is close to resolved: AI systems are tools, not authors, across every major scholarly integrity body’s current guidance. What remains genuinely in motion is the practice layer — disclosure format, granularity, verification, and cross-institutional consistency — and this is precisely where AI in research compliance functions will need to mature over the next several REF and grant-reporting cycles. Contributor role frameworks such as CRediT did not anticipate generative AI when devised in 2014, but their core discipline — mapping specific contributions to accountable agents — has turned out to be exactly the conceptual scaffolding institutions now reach for when drawing the line between tool and author. Research administrators who build that scaffolding into existing authorship and integrity workflows now will be far better placed than those who wait for a single global standard to arrive.
Leave a Reply