Why Forcing Scientists to Hand Over Every AI Prompt Could Backfire

A group of bioethicists and journal editors has launched a pointed challenge to one of the most popular ideas in the debate over artificial intelligence and scholarly publishing: the proposal that researchers who use large language models should be required to submit every prompt they typed, together with every output the models produced, as part of their manuscript submissions. Writing in the journal AI & Society, Brian D. Earp of the National University of Singapore, Udo Schüklenk of Queen’s University, Julian Savulescu of Oxford, and Sebastian Porsdam Mann argue that this policy, which they call Mandatory Full Inclusion, would fail at its central goal of policing authorship while inflicting collateral damage on academic freedom, equity, and the quality of scientific writing itself.

The concern driving calls for radical transparency is real. In a traditional workflow, named authors are presumed to have made a substantial intellectual contribution and to be able to answer for the integrity of their work. Large language models complicate that presumption because they can now help with brainstorming, structuring, rewriting, and even drafting entire manuscripts from a handful of instructions. Yet under prevailing guidelines, models cannot be listed as authors, since authorship is thought to require moral accountability that machines lack. The result is a gap: a researcher could, in principle, coax a minimally publishable paper out of an AI system, attach their name, and submit it without being detected, because the model will never object to being passed over.

The authors acknowledge that the problem may be worse than before the AI era. Prior to roughly 2023, anyone trying to claim credit for a paper they did not write had to contend with at least one other human collaborator, creating a natural disincentive. They cite a documented case, drawn from the secondary literature, of an academic producing articles on seemingly unrelated topics that may be almost entirely AI-generated. That author openly declares AI use and claims to have performed the editing, which technically satisfies current transparency standards, yet reviewers have no way of knowing whether one percent or one hundred percent of the text was machine-generated. The advocates of Mandatory Full Inclusion argue that requiring the raw receipts would let editors verify authorship claims directly, shifting the culture from blind trust in attestation statements toward verifiable trust based on evidence.

The new analysis does not dispute the motivation, but it dismantles the mechanism. The first and most obvious problem is unenforceability. A person unscrupulous enough to claim authorship on a largely AI-generated paper is, by hypothesis, someone willing to misrepresent their contribution, and such a person facing a choice between uploading materials that would immediately falsify their claim or simply not declaring AI use at all has strong incentives to choose the latter. AI detection software remains unreliable, and models can be prompted to scrub their own output of machine-like signatures. The likely outcome, the authors argue, is that unwarranted authorship claims would continue apace while honest declarations of AI use would decline, a result directly opposite to the policy’s intent.

Even for honest researchers, compliance may be technically infeasible. Serious scholarly work often involves consulting multiple models across many sessions, sometimes having the models critique each other’s responses to reduce the biases of any single training distribution. Built-in limits on context length, and the tendency of language models to underuse information in the middle of long interactions, give users further reasons to break work into separate exchanges. In collaborative projects where several co-authors each use one or more models over months, the resulting transcripts could run to hundreds or thousands of pages. Locating, exporting, redacting sensitive material, anonymizing for peer review, and assembling such a file could consume many hours of administrative labor that contributes nothing to the substance of the paper.

That burden creates what the authors call a transparency paradox, supported by recent empirical work. In a survey of twenty bioethics journals, only twelve papers out of more than one thousand published in 2024 referred to AI use or non-use, even though independent evidence suggests actual usage is many times higher. Experimental studies add that disclosed AI assistance can reduce perceived writing quality and trust relative to non-disclosure. A policy that attaches heavy documentation costs and potential stigma to every declaration would, the authors predict, push otherwise honest researchers toward omitting disclosure entirely, worsening the very under-reporting problem it aims to solve.

Perhaps the most novel objection concerns the Hawthorne effect, the well-documented tendency of people to change their behavior when they know they are being observed. If researchers know that every interaction with a model may be read by editors, reviewers, and readers, they may stop asking clarifying questions for fear of appearing under-informed, abandon creative back-and-forth for fear of seeming unserious, and avoid unconventional angles for fear of entertaining unpopular viewpoints. Evidence for this risk comes from a study of shared language model accounts, in which users reported that because conversational histories might be seen by others, they avoided personal or speculative topics, adopted alternative phrasings, deleted histories, or created separate accounts for certain queries. The authors argue that generative AI can serve as a private interlocutor during inquiry, a space for testing formulations, rehearsing objections, and exploring lines of thought that may later be abandoned. Mandatory surveillance of that space could shrink the exploratory thinking on which the best creative scholarship depends, and could produce a misleading kind of transparency: an artificially curated record shaped by the disclosure requirement itself rather than an accurate account of how ideas actually developed.

The equity costs fall unevenly. The authors identify four groups for whom language models can partly compensate for disadvantages that more privileged colleagues do not face: researchers whose first language is not English, who can use the tools to approach the fluency reviewers expect; researchers with conditions affecting writing, such as dyslexia, ADHD, chronic fatigue, repetitive strain injury, multiple sclerosis, or motor impairments; researchers with heavy caregiving responsibilities, who are disproportionately women and who may benefit from drafting in short bursts; and scholars at under-resourced institutions, particularly in the Global South, who often lack the postdocs, editorial assistants, and writing support available at well-funded universities. Suppressing AI use through surveillance-heavy documentation requirements could therefore widen existing disparities in who gets published, even if institutions must simultaneously pursue the structural reforms that address the root causes of those disparities.

The policy also asks the impossible of the people who would have to do the checking. Peer reviewers are selected for expertise in a manuscript’s subject matter, not for the ability to judge from a prompt transcript how much of a text reflects substantive human contribution versus generic machine generation, a skill no current training program cultivates. Reviewer recruitment is already a serious bottleneck, especially in the humanities and qualitative social sciences, where editors may struggle to find even one suitable expert per submission. The predictable result is that supplemental files would go unread or receive only superficial assessment, defeating the policy’s purpose. Moreover, the authors note, academic publishing has never demanded this level of documentation of anyone: researchers are not required to upload minutes of every meeting, every tracked-changes draft, or records of which paragraph belongs to which co-author, because the costs and chilling effects are judged to outweigh the benefits. Singling out AI use for demands far exceeding those imposed on human collaborators is inconsistent with the field’s established, and well-justified, balance between transparency and other goods.

In place of Mandatory Full Inclusion, the authors propose a proportionate framework. First, routine declaration of AI use in all manuscripts, whether or not AI was used, to normalize transparency; one proposed scheme offers three simple options: no generative AI use, AI limited to copyediting human-drafted text, or substantive AI use described narratively. Second, targeted methods-level disclosure, including prompt logs where necessary, when AI bears on scientific validity or reproducibility, such as in data analysis, code generation, or structured literature synthesis, consistent with emerging reporting checklists. Third, voluntary submission of prompts and outputs for authors who wish to document their own contribution, potentially through summary-level methods such as the SLICE approach, in which independent assessors estimate the percentage of human and AI input at each stage of writing without exposing raw transcripts. Finally, journals should retain and enforce the requirement that anyone claiming authorship explain how they meet the relevant criteria, through ICMJE-style attestations or CRediT statements, and treat false attestations as misconduct. The authors concede that a trust-plus-accountability system is imperfect and that some will exploit it, but they conclude that mandatory full prompt disclosure would not fix that problem and would instead introduce surveillance harms, discourage beneficial uses of AI, and disadvantage the very scholars who need these tools most, ultimately doing more harm than good.

Subject of Research: Ethics of mandatory disclosure of AI prompts in scholarly publishing

Article Title: Against mandatory prompt disclosure in AI-assisted scholarship

Article References: Earp, B. D., Schüklenk, U., Savulescu, J., & Porsdam Mann, S. (2026). Against mandatory prompt disclosure in AI-assisted scholarship. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03326-w

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03326-w

Keywords: large language models, academic publishing, research integrity, authorship, prompt disclosure, transparency, Hawthorne effect, peer review, AI ethics, scholarly writing, equity, research policy