From ChatGPT to Genuine Co-Author: How to Use AI in Scientific Publications Without Losing Authorship

There is a conversation happening in every research group that has not yet found its way into any institutional policy document, any supervisor-student meeting, or any formal guidance from the funding bodies that pay for the science. It goes roughly like this: someone used ChatGPT to draft the introduction. Someone else ran the literature review through Elicit. The discussion section was written in English by a researcher whose first language is Georgian, with significant AI assistance to make it readable. None of this was disclosed in the manuscript because nobody was entirely sure what they were required to disclose, and the prevailing assumption was that everyone was doing the same thing, and nobody was really checking.

This assumption is now materially wrong on both counts. Not everyone is doing it with the same intent or the same care. And people are checking — with increasingly accurate tools, applied increasingly routinely, at the editorial offices of every journal that matters.

The question is not whether to use AI in scientific writing. That question has been settled by the behaviour of the research community itself. The question is how to use it in ways that serve the science, protect the researcher's intellectual ownership of the work, satisfy the disclosure requirements that journals are now enforcing, and produce manuscripts that neither misrepresent the nature of the contribution nor expose co-authors to consequences they did not knowingly accept.

The Authorship Question: Why AI Cannot Be a Co-Author

The consensus across every major editorial body is complete and unambiguous. Any AI tool cannot fulfil one of the basic criteria of authorship according to ICMJE: taking responsibility for the accuracy, integrity, and originality of a paper's contents. Therefore, no AI tool, such as ChatGPT, can be listed as an author, even if it has been extensively used in the study or paper.

The ICMJE outlines four essential criteria for authorship: substantial contributions to the conception or design of the work; drafting or revising it critically; final approval of the version to be published; and agreement to be accountable for all aspects of the work. ChatGPT, as an AI system, lacks cognitive intentionality, legal accountability, and the ability to engage in ethical reasoning. It cannot provide critical analysis, endorse intellectual content, or assume responsibility for published material.

A central principle of the 2026 ICMJE update is that AI tools or assisted technologies cannot be listed as authors. Because AI systems cannot assume responsibility for the accuracy, integrity, or originality of scholarly work, authorship remains exclusively a human attribution.

This is not a bureaucratic technicality. It is the articulation of something philosophically important about what scientific authorship means — and understanding it clearly is the foundation for using AI in ways that do not erode it.

Authorship in science is not a credit for producing text. It is an accountability claim. When a researcher's name appears on a paper, they are asserting that they understand the work, stand behind its integrity, and can defend its claims under challenge. An AI system cannot make that assertion in any meaningful sense — it cannot be summoned before a journal editor, it cannot respond to a post-publication critique, it cannot retract a finding when new evidence emerges, and it cannot be held responsible for a fabricated citation. The human authors can. That is why they are the authors, and why that status cannot be shared with a tool regardless of how much text that tool generated.

The practical implication is this: if AI writes a significant portion of your manuscript, you are not thereby absolved of any part of your authorship responsibility. You are, if anything, more responsible — because you have introduced a layer of text generation that you did not fully control, and you must verify everything it produced before your name goes on it.

The Disclosure Landscape: What Different Publishers Actually Require

The disclosure requirements across major publishers are converging on a common framework while retaining enough variation to make journal-specific checking essential before submission.

Wiley requires disclosure of AI when used to generate substantial text or restructure arguments: in the Acknowledgements when AI assists with manuscript drafting, editing, translation, or formatting; in the Methods when AI is used in research methodology, data collection or analysis, or literature review; and in figure captions when AI generates or edits any visual content. Simple grammar correction, formatting, and word choice suggestions typically do not require disclosure.

Publishers permit AI as a supportive tool for tasks like improving language, brainstorming, or summarising literature — but not for replacing core intellectual work like analysis or drawing conclusions. The "human-in-the-loop" is a non-negotiable requirement, mandating that authors critically review, edit, and take responsibility for all AI-generated output. The failure to exercise intellectual ownership and critical evaluation — not the use of the tool itself — is what constitutes academic misconduct.

The practical summary of where the line sits is as follows: using AI to improve how you express ideas you have already formed is generally permitted with appropriate disclosure. Using AI to form the ideas — to generate the analysis, to draw the conclusions, to synthesise the literature in ways that substitute for the researcher's own critical engagement with the field — is the territory where the authorship claim becomes problematic. The distinction is not always crisp in practice, which is precisely why disclosure and verification are the non-negotiable minimum.

The Legitimate Uses: Where AI Genuinely Helps

There is a meaningful difference between the uses of AI that genuinely support scientific writing while leaving intellectual ownership entirely with the researcher, and those that begin to substitute for the researcher's own thinking. The former category is substantial, practically valuable, and entirely compatible with honest disclosure.

• Language editing and fluency improvement

  The use case with the clearest ethical standing and the greatest practical value for the largest number of researchers. A scientist in Kyiv or Almaty who has designed and conducted original experiments, analysed the results with genuine intellectual rigour, and drafted a manuscript in English that is scientifically accurate but stylistically imperfect is not compromising their authorship by using AI to make that draft more fluent. They are using a tool the way they would use a dictionary — to express more accurately what they already know.

• Literature search and organisation

  Tools such as Elicit, Consensus, and Semantic Scholar can identify relevant papers across large bodies of literature, generate structured summaries, and organise sources by relevance — tasks that previously required weeks and are now achievable in hours. The intellectual work of evaluating what the literature means, identifying gaps, and positioning a contribution within it remains entirely the researcher's own. The tool accelerated the information gathering; the judgement is the researcher's.

• Structural drafting

  Using AI to generate a first outline of how a manuscript might be organised, which the researcher then populates with their own content, falls in acceptable territory provided the intellectual substance of every section originates with the researcher. An AI-generated outline that the researcher rejects, revises, or uses as a starting point for their own structure is a thinking aid, not a ghost-writing service.

• Code generation and computational assistance

  Data analysis pipelines, figure formatting, and statistical computation fall under the same principles as other methodological tool use — disclosed in the Methods section, outputs verified by a researcher who understands what the code is doing, and the intellectual interpretation of the results must remain the researcher's own.

• Translation between languages

  Particularly for researchers working across Russian, Georgian, Uzbek, or other CIS languages and English — a legitimate and practical use, provided the translated content is verified by a researcher with sufficient command of both languages to confirm that the scientific meaning has been accurately preserved.

Where Authorship Erodes: The Uses That Create Risk

The uses of AI that genuinely compromise authorship are not always the obvious ones. The outright generation of a manuscript from scratch is clearly problematic and increasingly easily detected. But the erosion of authorship happens more often through accumulation of smaller substitutions — each individually defensible, collectively constituting a paper whose intellectual substance the named authors did not actually produce.

The signal that authorship is eroding is not how much text the AI generated. It is whether the researcher who submits the paper can defend every claim in it without reference to what the AI produced. If a researcher used AI to summarise a body of literature in the introduction and has not themselves read the primary sources the AI cited, they cannot defend the framing of the introduction under challenge — and that inability is a genuine authorship failure, not a technical one.

The specific failure modes that are producing retractions and post-publication challenges in 2025 and 2026 are:

1. Hallucinated references that were not verified

   An AI tool asked to support a claim will produce plausible-looking citations. If those citations are not checked against their actual sources, the paper may contain references to papers that do not exist, or that do not say what they are claimed to say. Either is grounds for retraction. This is the single most preventable category of AI-related retraction — and the single most common.

2. AI-generated text that misrepresents your results

   A researcher who asks AI to write their discussion section based on their results, and who does not carefully verify that the discussion accurately represents those results rather than a plausible-sounding generalisation of them, has introduced a potential misrepresentation of their own work. The danger here is subtle: the AI output is plausible, internally consistent, and may even be eloquent — and yet wrong about what the research actually showed.

3. Undisclosed AI use discovered after publication

   The detection infrastructure described in the previous article in this series is real, improving, and being deployed by major journals. A paper with substantial undisclosed AI involvement that is subsequently identified is a retraction risk regardless of the quality of the underlying science. The longer between submission and detection, the worse the institutional and reputational consequences.

A Framework for Thinking About AI Use in Your Manuscript

Rather than a checklist — which exists in the previous article in this series — this article offers a framework of three questions that, answered honestly before submission, will locate any specific AI use in the correct position relative to authorship and disclosure requirements.

A research team that can answer all three questions honestly and affirmatively has used AI in ways that are compatible with genuine authorship, honest disclosure, and the integrity standards the community is now enforcing. A team that cannot is carrying a risk that will, increasingly, materialise.

The Authorship That Remains Yours

There is a version of this conversation that treats AI as a threat to scientific authorship — a technology that will gradually hollow out the intellectual content of research papers until nothing genuinely human remains. That framing misses something important.

The things that AI cannot do — form a hypothesis, design an experiment, evaluate whether a result is surprising, judge whether a conclusion is warranted by the evidence, take responsibility when the answer is wrong — are precisely the things that make scientific authorship meaningful. They are not incidental to the scientific enterprise. They are the scientific enterprise.

The researcher who uses AI to express their ideas more clearly, to find literature they would otherwise have missed, to run their analysis pipeline faster, and to check their English grammar has not given anything away. They have the same ideas, the same intellectual responsibility, and the same claim to authorship they had before the tool existed. They have simply done more with the time they have — which is, after all, the point of every tool science has ever adopted.

What they owe their co-authors, their readers, their journals, and themselves is honesty about what the tool did and what they did. That honesty is not a bureaucratic requirement. It is the condition under which the authorship claim means anything at all.