Who Is Responsible When AI Helps to Write Science?

Posted on 2 September, 2026 by Mehrdad Fathi

Generative AI is now embedded in the daily workflow of research. It summarizes literature, drafts sections, sharpens arguments, and increasingly shapes the conceptual framing of papers. Yet publishing norms have not kept pace. The question that remains unanswered is not whether AI is being used — it clearly is — but who is accountable for what it produces.

Authorship was never just a credit mechanism. It is an accountability structure: a signal to the community that a named scholar stands behind the work, can defend it, and can be held responsible for its integrity. When genAI mediates part of that process, the existing structure begins to break down in ways that matter.

The Accountability Gap

Traditional research teams divide labor along lines that have always been unequal. Junior researchers collect data and mine the literature. Senior researchers conceptualize, synthesize, and revise. Each person can explain their contribution, respond to criticism, and be corrected. GenAI can do none of those things. It produces outputs but bears no responsibility for them.

The Contributor Roles Taxonomy — CRediT — was designed to make human contributions legible beyond simple author order. It does that well. But it was built for people, not for human–AI workflows. As a result, current authorship statements can no longer reliably tell readers who did what, how AI was involved, or who is accountable for the result.

The problem extends beyond writing. AI is beginning to shape editorial workflows too: reviewer selection, interpretation of peer review reports, the language of editorial decisions. These processes determine what becomes scholarship and who earns authority in the field — yet they remain largely invisible under current disclosure practices.

The Invisible Labor Question

The acknowledgements section of a paper is not mere courtesy. It is a graded economy of attribution, and it reflects what the field values. Some contributions rise to authorship; others are downgraded to “helpful comments”; others vanish entirely. This hierarchy has always been imperfect. GenAI makes its imperfections more visible — and more consequential.

Notably, the human labor embedded in genAI systems themselves rarely surfaces in this economy: the workers who label training data, evaluate model outputs, maintain platforms, and produce the texts on which models are built. As long as current disclosure norms ignore AI mediation, that labor stays invisible.

This visibility problem deserves more than compliance-level attention. If AI-assisted literature synthesis were attributed through authorship, the same question would naturally arise: why does equivalent work performed by a human research assistant not earn them the same recognition? The advent of genAI creates pressure to examine the attribution economy itself — not merely to add AI as a new line item within it.

A Practical Path Forward: CRediT-AI

The article proposes linking AI disclosure directly to existing CRediT roles rather than appending a generic end-of-paper statement. A generic statement says that AI was used. A CRediT-AI statement would describe how it mediated each specific human contribution.

The structure is straightforward. Each relevant CRediT role would carry a brief description of AI involvement: “AI-assisted ideation and framing of alternatives” for conceptualization; “AI-assisted candidate prose generation and structural editing” for writing. Roles with no AI involvement would note “none.” Detailed prompts, workflows, or model parameters would sit in methods sections or supplementary material where reproducibility demands them.

Crucially, the named human author would still hold the CRediT role. The AI disclosure would describe how that contribution was accomplished — not replace the person accountable for it. A verification clause would accompany each disclosure: an explicit statement identifying who reviewed the AI-mediated output, whether someone validated the references used in a literature synthesis or checked the accuracy of AI-drafted claims.

The framework also acknowledges structural inequity. Access to genAI tools is uneven across institutions and geographies, and that asymmetry should surface in disclosure rather than be hidden behind neutral-sounding methodology language.

Why This Matters Beyond Publishing

The argument here goes further than process reform. Academic publishing is the infrastructure through which scholarly authority is established and contested. Decisions about AI’s role in that infrastructure — in submission screening, peer review, grant evaluation, and editorial workflows — should not be driven by convenience or cost savings alone. They warrant public debate.

GenAI has made the existing conditions of scholarly production visible in ways they were not before. The authors treat that visibility as an opening for negotiation, not a problem to be smoothed over. The CRediT-AI framework they propose is one practical intervention among possible others — a starting point for accountability to catch up with practice, with practice, not a final settlement.

For researchers and institutions navigating AI adoption, that framing is the more important takeaway. The tools ase. The question of who is responsible for what they produce is still open — and the field gets to decide how to answer it.

Source: Adapted from Nature 657, 34–36 (2026). doi: 10.1038/d41586-026-02686-z

References:

1. Rees, G. & Wilsdon, J. Nature 652, 1119–1121 (2026).

2. Earp, B. D., Guernon, A.-S. & Porsdam Mann, S. Nature Reviews Bioengineering 4, 479–480 (2026).