CASRAI Dictionary

Tag: authorship

  • Disclosing generative AI use in research: what to declare and where

    Two or three years ago, declaring the use of a generative AI tool in a manuscript was an unusual courtesy. Today it is a baseline expectation, written into the author instructions of most major publishers and the recommendations of the bodies that set publishing norms. Yet the question authors most often ask is disarmingly practical: what exactly do I have to declare, and where does the declaration go? This article sets out a clear answer, drawing on the vocabulary being developed in the generative AI use and disclosure domain.

    The two settled principles

    Underneath the variation between publishers, two principles have hardened into near-consensus, and they are the right place to start.

    The first is that a generative AI system cannot be an author. The ICMJE recommendations, and parallel statements from COPE, Nature, Science, and the major university presses, are explicit on this point: authorship entails accountability for the work, and a tool cannot be accountable. AI use is therefore disclosed as a method or a tool, never as a contributor on the author line. This connects directly to the broader account of authorship as a matter of responsibility, not merely of having touched the text.

    The second is that the human authors remain fully responsible for everything the manuscript asserts, including anything an AI system produced. A fabricated citation, a misstated statistic, or a plausible-but-wrong sentence is the authors’ error regardless of which tool generated it. Disclosure does not transfer responsibility; it makes the workflow transparent so that responsibility can be located.

    What counts as disclosable use

    The harder question is the threshold. Not every interaction with a computational tool is a disclosable use of generative AI, and policies generally exempt the trivial. The useful distinction is whether the tool produced novel content that materially shaped the published work.

    • AI-assisted writing — where a generative system drafted, restructured, summarised, or substantively edited text whose output shaped the published wording — is disclosable. A generative AI tool is, in the working definition, a system that produces novel text, code, image, or other media from a prompt, typically using a large neural network.
    • AI-assisted analysis — using a model to perform or shape a data-analysis step, including exploratory analysis or hypothesis generation — is disclosable as part of the methods.
    • AI-generated code that forms part of the research, and AI-generated images in a manuscript, are disclosable, the latter often under stricter rules because of the integrity risks around figures.

    By contrast, most policies define an AI use exempt category for tools that do not produce novel content: a spell-checker, a grammar corrector, a reference manager, or basic translation of the author’s own words. Author-written text whose grammar was tidied by an AI checker is not, in this sense, AI-assisted writing. The line is not always crisp — substantive rewriting shades into drafting — and when in doubt the safe practice is to disclose.

    Where the declaration belongs

    Knowing what to declare is half the problem; the other half is placement, and here practice has converged on a small set of locations.

    The dominant convention is a dedicated AI use disclosure statement in the manuscript: a short declaration that names the system, says where in the workflow it was used, and indicates the extent of that use. “Which tool, where, and how much” is the durable shape of a good statement. Many journals place this in the methods section when the use was analytical, and in a distinct acknowledgements-adjacent statement when the use was in writing.

    A useful test for a disclosure statement: a reader should be able to tell, from the statement alone, which parts of the work involved a generative system and what the authors did to verify its output. A generic line that an AI tool was “used to improve readability” fails this test; it names neither the tool nor the boundary of its use.

    Two adjacent practices strengthen the statement. The first is recording a model selection rationale and, where relevant, the prompt engineering that produced reliable outputs — material that belongs in supplementary methods for analytical uses, because it bears on reproducibility. The second is naming the AI tool provider at the organisational level, so that the disclosure points at an identifiable system rather than a generic category.

    Why structured disclosure, not just prose

    A free-text paragraph at the end of a manuscript is where most disclosures live today, and it is better than nothing. But prose disclosure has the same weakness that prose contribution statements have: it does not travel as data. A structured representation — naming the tool, the workflow stage, the extent, and the verification step as discrete, machine-readable fields — lets downstream systems index, audit, and aggregate AI use across the literature. That is the difference between a sentence a human must read and a record a system can act on, and it is the gap a controlled vocabulary is meant to close. The parallel with structured contribution metadata in CRediT is exact: a settled human-readable form, waiting on consistent machine-readable plumbing.

    The role for shared vocabulary

    Publishers’ AI policies differ in wording, in threshold, and in placement, which means a disclosure written for one journal does not necessarily mean the same thing when read by another system. What is missing is not more policy — the principles are settled — but a shared definitional layer: agreed terms for AI-assisted writing, AI-assisted analysis, exempt category, and the rest, so that a disclosure carries the same meaning wherever it is read. Supplying that layer, federating to ICMJE and COPE for the normative content rather than inventing it, is the convening role the CASRAI dictionary is built for. The practical guidance for authors lives at AI disclosure for authors.

    What to do now

    For authors: disclose any use that produced novel content shaping the work, name the tool and the workflow stage, and state that you verified the output. For editors: specify where the statement goes and ask for structured fields, not just a paragraph. For standards work: prioritise shared definitions of the disclosable categories and the exempt threshold, so disclosures mean the same thing across venues.

    Related reading

  • AI agents and autonomous research: attribution and accountability

    For most of the history of science, the tools of research — however sophisticated — did the bidding of the people using them. A telescope or a statistical package extended human capability but did not decide what to investigate. That assumption is now being tested. AI agents capable of a degree of autonomy are beginning to appear in research: systems that can generate hypotheses, design experiments, and in some cases run them through automated laboratory equipment, iterating with limited human intervention. Autonomous experimentation of this kind raises a question scholarship was never built to answer: when an AI system materially contributes to a discovery, how should that contribution be attributed, and who is accountable for it? This article examines those questions, drawing on the AI and ML research-outputs domain of the CASRAI Dictionary.

    What autonomous research looks like

    The systems in question share a common feature: they make consequential choices in the research process rather than merely executing instructions. An AI agent might propose which compounds to test next, design the sequence of experiments, control the apparatus that performs them, and analyse the results to decide what to try next — a loop that can continue with the human supervisor stepping in only occasionally. The appeal is obvious: such systems can explore vast spaces of possibility far faster than people, accelerating discovery from materials science to drug development. But the autonomy that makes them powerful is what unsettles the established account of who does research and who answers for it. The agent is no longer just a tool; it is participating in the intellectual work. That shift forces the questions of attribution and accountability.

    Why an AI cannot be an author

    The clearest and most settled point in this debate is also the most important: an AI system cannot be an author of a research work. This is not technophobia or an arbitrary rule; it follows directly from what authorship means. Authorship carries accountability. An author is someone who can take responsibility for the integrity of the work, vouch for its honesty, defend it when questioned, and be answerable if it proves flawed or fraudulent. An AI system can do none of these things; it cannot be held responsible or called to account. The major editorial and integrity bodies have converged firmly on this position: AI tools, however capable, cannot meet the criteria for authorship, because the defining quality of an author — answerability — is one a machine cannot possess. The principles of authorship rest on responsibility, and responsibility is irreducibly human.

    Accountability stays with people

    If the AI cannot be accountable, accountability does not vanish — it remains with the humans involved. The researchers who deploy an autonomous system, decide to use its outputs, design the study it operates within and interpret and publish the results are responsible for that work, including for the AI’s contributions to it. This has a sharp consequence: a researcher cannot disclaim responsibility for an error or fabrication by pointing to the machine. If an AI agent generates a flawed hypothesis and a researcher publishes it as sound, the failure is the researcher’s, because the duty to verify and stand behind the work was theirs. Far from diluting human responsibility, autonomous systems concentrate it: the more capable the tool, the more important the human judgement about whether and how to trust it. Autonomy in the tool does not mean autonomy from accountability for the people.

    Disclosure and the provenance of AI contributions

    If an AI agent cannot be credited as an author but did genuinely contribute, the honest course is to describe what it did transparently. This is a matter of disclosure and provenance rather than authorship. A research report should be clear about the role an autonomous system played — which hypotheses it generated, which experiments it designed, which analyses it performed — so readers can understand how the work was produced and judge it accordingly. Recording the provenance of AI contributions serves several ends at once:

    • Transparency. Readers and reviewers can see where machine judgement entered the work and weigh it appropriately.
    • Reproducibility. Knowing which system was used, and how, is part of being able to reproduce the result.
    • Accountability. Clear provenance makes plain which choices were the system’s and which the researchers’, keeping responsibility traceable.

    Disclosure does not credit the machine; it documents it — an entirely different and appropriate act.

    The limits of CRediT

    It is natural to ask whether a contribution taxonomy could simply add the AI as a contributor. Here it is worth being precise about what the CRediT taxonomy is for. CRediT describes the contributions of people to a research work; it is a vocabulary for human roles, anchored in the assumption that contributors are accountable agents. An autonomous system is not a contributor in that sense, because it cannot bear the responsibility contributorship implies. The right place for AI involvement is therefore not the contributor list but the methods and disclosure sections, where its use can be described as part of how the work was done. What CRediT continues to do well is capture the human contributions around the AI — the conceptualisation, methodology, investigation and interpretation that remain human even when a machine assists. The taxonomy’s limits here are not a defect; they reflect the correct distinction between a tool that is used and a person who is answerable.

    A consistent vocabulary for a changing landscape

    As autonomous systems become more common, describing their involvement consistently — what was used, for what, and where human responsibility sat — will matter increasingly across journals and institutions. That consistency is what the CASRAI Dictionary works towards: a shared vocabulary so a statement about how AI contributed to a piece of research, and who is accountable for it, is understood the same way wherever it is recorded. AI agents may transform the pace of discovery; the durable principles — that authorship means accountability, that responsibility stays with people, and that AI contributions are disclosed rather than credited — are what keep research trustworthy as the tools grow more powerful.