Who is responsible when AI helps to write science? As generative AI becomes embedded in research and publishing, the deeper challenge is preserving accountability for scholarly work.

“Generative artificial intelligence (genAI) tools are increasingly being used in science to search and summarize literature, produce ideas, draft responses to reviewers and improve the style of written text. This raises a question: what counts as human authorship when some of a researcher’s work is mediated by AI? And how much of this intellectual ‘work’ should be attributed to an AI model or its developers?

 

Authorship in science is an earned status: it recognizes a scholar’s contribution to a paper and signals to the research community that they are a competent and responsible scientist. What is at stake in potentially recognizing AI as an author is not simply who gets credit for a text, but how scholarly writing is deemed accountable and trustworthy.

 

AI, peer review and the human activity of science

 

This problem pre-dates genAI. Early modern authorship was not a natural or fixed status but a practical achievement, established through conventions. Isaac Newton’s authority, for instance, rested not only on scientific discovery, but also on his ability to navigate the institutions such as the Royal Society and publication practices through which scholarly credit was assigned. New technologies have repeatedly reshaped these arrangements.

 

Ultimately, current decisions regarding AI’s role in academic publishing have a bearing on the future of scholarship itself1. If AI is increasingly used to screen submissions, assist with peer review, create grant proposals, evaluate research outputs and shape publishing workflows, then decisions about its use should not be driven solely by convenience or cost saving. More explicit public debate is needed.

 

Here, I examine some of the challenges for scholarship surfaced by the advent of AI and outline steps forwards.

AI-mediated authorship

 

Academic writing has long been organized through differentiated, hierarchical labour. In a research team, junior members might search the literature and collect data. More experienced scientists can develop the conceptual framing of a paper and draft its sections, and senior researchers often supervise, revise and approve the final version of the work.

 

When authors use genAI tools to do part of their work, it’s not equivalent to bringing in a human contributor. A graduate student or research assistant can explain their contribution, respond to criticism, learn from correction and be held accountable for the integrity and accuracy of their work. GenAI is different: it cannot justify or take responsibility for what it produces.

 

The uncritical adoption of AI in science is alarming — we urgently need guard rails

 

Use of genAI involves a set of technologically mediated practices: writing prompts, generating outputs, summarizing information, classifying content and checking results. These actions fall outside current authorship categories — and thus authorship cannot reveal who did what, how GenAI was involved or how credit and responsibility should be assigned. Without clear rules for AI-mediated work, authorship becomes harder to account for.

 

Existing authorship-role taxonomies, such as CRediT — the Contributor Roles Taxonomy, now used by several publishers, which specifies the roles researchers had in producing a paper — are both necessary and insufficient. They remain valuable because they break down scholarly labour into specific roles, going beyond the misconception that author order alone captures contributions. But CRediT was built for human contributors. Human–genAI writing upends it; current reporting structures barely register AI mediation.

 

Beyond the writing team, the politics of authorship extends through the editorial and review infrastructures that manuscripts must go through to be considered scholarship. AI has started playing a part in this arena, too. Reviewer selection, editorial interpretation of reports and the language of decisions — all of which can be influenced by AI use — shape the fate of the text and the authority of its eventual authors.

 

AI scientists are changing research — institutions, funders and publishers must respond

 

The politics of authorship is particularly evident in the acknowledgements section. These statements are not just courtesies. They are part of a graded economy of attribution through which scholarly labour is valued or left invisible. Some work becomes authorship, some gets downgraded to ‘helpful comments’ or ‘assistance’ and some disappears.

 

In this section, research assistants, students, reviewers and colleagues encountered at workshops, conferences or dinner tables might be mentioned, named and thanked — or left out. Such omissions extend to the largely invisible human labour that makes genAI systems work: data labelling, model evaluation, engineering work, platform maintenance and the production of the texts on which the models are trained.

 

This is why the current debate should be about more than solving the problem of how to fit genAI into an existing graded economy of attribution. It is an opportunity to question that economy itself. By necessitating new categories of authorship practices in existing frameworks, genAI can create an opening to destabilize some hierarchies in academic work, rather than merely adding a new tool to the workflow: if AI-assisted literature synthesis is recognized through authorship, why would the same work, when performed by a human research assistant, not earn them authorship as well?

New credit framework

 

Experiments in AI-assisted scholarship are beginning to test alternative ways of organizing authorship and responsibility2.

 

A useful example is a paper on constitutional AI (a training method in which a model is given a set of human-written ethical rules), published this year (see go.nature.com/67ysb9). Andrew Maynard, who researches advanced technology transitions at Arizona State University in Tempe, reports that Claude Opus 4.6 researched and wrote the paper with little to no oversight, but Maynard remained the named author and accepted responsibility for the work. The human–AI division of labour is part of the paper’s argument and is documented in a postscript. The example is informative, but shows the limits of bespoke disclosure: a detailed narrative can explain one paper well, but it is not standardized, comparable across papers or linked to established contributor roles.

 

Instead of adding a generic AI-use statement at the end of a paper, journals could link AI disclosure to existing reporting of contributions. A generic statement says that AI was used; a CRediT-AI statement would indicate the human contribution that it mediated and for what purpose (see ‘Example CRediT-AI structure’).

 

The aim would be to describe, briefly, how each human contribution was made. Each relevant CRediT role could be paired with a short description of AI mediation. For example, ‘AI-assisted ideation and framing alternatives’ for conceptualization, or ‘AI-assisted candidate prose generation and restructuring’ for drafting. Roles with no AI involvement could simply say ‘none’. Only AI use that materially shaped a contribution would need to be recorded at this level; detailed prompts, workflows or model settings could remain in the methods section or supplementary information when they are important for interpreting or reproducing the work. The human author would still hold the CRediT role; the AI statement would describe how that contribution was accomplished.

 

Any such framework would need a verification and accountability statement, affirmed by the authors. It would identify who reviewed AI-mediated contributions. When AI directly shaped the output associated with a CRediT role, the statement could indicate who was responsible for verifying that contribution — for example, checking references used in literature synthesis or validating code.

 

Thus, a CRediT-AI reporting structure could move beyond technical reporting towards who is using AI, why it is needed and who is accountable. An author might also note that genAI was used mainly for English-language reformulation, was accessed through a paid institutional subscription and entailed computational and energy costs that could not be independently quantified. Such disclosure would not capture every inequality associated with genAI use, but it would keep authorship connected to the wider conditions under which AI-mediated scholarship is accomplished.

 

The CRediT-AI framework proposed here is not a final settlement; it is one possible practical intervention. It would not resolve the question of authorship once and for all. But it could show how human responsibility is being accomplished in AI-mediated scholarship. And that such disclosure is possible.

 

GenAI has shed fresh light on authorship and the infrastructures through which written work becomes scholarship. That visibility should not be seen as a problem to be suppressed, but as an opening for negotiation over the future conditions of scholarly life.

GenAI (OpenAI ChatGPT 5.4) was used as a dialogical writing and analytical aid during the development of the piece, including for idea generation, structural iteration, prose generation, summarization, condensation and stylistic revision. It also assisted in synthesizing relationships between the cited papers and the author’s uploaded manuscripts. GenAI (Anthropic Claude 4.2) was used to review the article. Review comments have been addressed during dialogical revision completed with ChatGPT." [1]

 

1. Who is responsible when AI helps to write science? As generative AI becomes embedded in research and publishing, the deeper challenge is preserving accountability for scholarly work. By Robert Braun. Nature 657, 34-36 (2026)