CASRAI Dictionary

Tag: AI disclosure

  • CRediT for AI-generated content: where the line is

    The ICMJE 2023 position is settled: artificial-intelligence systems cannot be authors. The follow-on question that journals and authors continue to negotiate is how to represent, in a contributorship statement, the human work that goes into producing AI-assisted content. When a co-author prompts an LLM to draft a section, verifies the output, edits it, and stands behind it, which CRediT role describes their contribution? This post proposes a working line.

    The shape of the question

    Three scenarios make the question concrete.

    Scenario one: an author uses an LLM to polish prose in a draft they wrote. The intellectual content is theirs; the language is partly the model’s. The CRediT role is straightforwardly Writing – original draft for the author.

    Scenario two: an author uses an LLM to draft a first version of a section, which they then heavily revise. The first draft is the model’s; the final draft is the author’s, but the model substantively shaped what the final draft says. The CRediT role is still Writing – original draft for the author, but the contribution is meaningfully different from scenario one.

    Scenario three: an author uses an LLM to propose a study design, which they then refine. The intellectual content of the methodology was partly the model’s. The CRediT role is Methodology for the author, but again the contribution is meaningfully different from the unaided version.

    In all three, the human author is the role-holder; the model is not a co-author. What is different across the scenarios is the magnitude and the character of the human contribution. CRediT, as currently constituted, does not distinguish these.

    The working line

    Our proposed line is the verification-and-responsibility threshold. A human contributor who has substantively verified the AI-generated content, taken responsibility for it, and is prepared to defend it in correspondence or post-publication discussion is properly credited with the relevant CRediT role. The role describes what they contributed to the paper, which includes verification work even if the first-draft work was the model’s.

    The line shifts where the human contribution is insubstantial — a contributor who pasted a prompt, accepted the output without verification, and added their name to the paper has not discharged the role and should not be credited. This is the same line that has always applied to non-AI cases (a co-author who did not contribute should not be credited; gift authorship is a well-recognised failure mode).

    The line is therefore not about AI use per se; it is about whether the human contribution clears the substantive-contribution threshold. AI use does not displace the threshold; it changes what discharging the role looks like in practice.

    Disclosure runs parallel

    The disclosure of AI use is a separate question, addressed via publisher-mandated AI disclosure declarations. The disclosure says what tools were used and for what; the CRediT statement says who contributed what to the paper. The two run parallel and are both required by most major publishers in 2026. The CASRAI AI disclosure for authors guide walks through the publisher-by-publisher requirements.

    Implications for specific roles

    Writing – original draft

    The most common case. A human author whose draft was AI-assisted is properly credited with Writing – original draft if they verified the content, took responsibility for it, and produced the version that is the paper. The disclosure declaration says the AI was used; the CRediT statement names the human as the writer-of-record.

    Methodology and Formal analysis

    More delicate. If an AI-assisted statistical-discovery tool proposed a method or an analytic approach, the human contributor’s role is partly verification (was the proposal sound?) and partly extension (refining the proposal into the actual method). The CRediT role is still Methodology and/or Formal analysis for the human, but the verification dimension is foregrounded. If the human did not verify — accepted the AI proposal without independent assessment — the contribution is weaker and may not clear the role threshold.

    Investigation

    A subtle case. AI-assisted data extraction (e.g., from imaging, from medical records, from text corpora) involves a human contribution that runs from setup through verification to interpretation. Investigation includes the data-gathering activity; an AI-assisted version still has a human Investigation lead, who is responsible for the setup, the verification of extracted data, and the handling of errors.

    Validation

    Perhaps the most directly affected. Where AI tools are used for cross-checking, sensitivity analyses, or reproduction of results, the human Validation contributor is responsible for setting up the validation, interpreting its results, and acting on discrepancies. The AI does the mechanics; the human does the judgement.

    Visualization

    AI-assisted figure generation is increasingly common. The human Visualization contributor is responsible for the figure-design decisions, for verifying that the AI-generated figure accurately represents the data, and for the final version that appears in the paper. Where the AI generated an image that the human did not substantively verify, the threshold may not be cleared.

    The role-as-recognition trap

    A failure mode to flag explicitly. The temptation, when AI did most of the actual production work, is to inflate the human contributor’s role assignment to compensate. “The AI wrote the draft, but I prompted it, so I should still be Lead on Writing – original draft.” This is a misreading. The CRediT role is a description of contribution; if the human contribution was “prompted and accepted”, that is a smaller contribution than “drafted, verified, revised, took responsibility.” Calling both “Lead” obscures the difference.

    The remedy is the degree-of-contribution qualifier. A human contributor whose AI-assisted contribution was substantial may be Lead; one whose contribution was lighter may be Supporting. The qualifier discipline forces an honest assessment of magnitude.

    Where this leaves the AI-assistance-role question

    We have argued elsewhere that a 15th CRediT role explicitly for AI assistance is worth considering. The argument from this post is partly orthogonal: the existing 14 roles can accommodate AI-assisted work if the verification-and-responsibility threshold is honoured and the qualifier is used honestly. The case for a 15th role rests on whether the structured disclosure-of-AI-use is better placed inside the contributorship statement or outside it. Reasonable people disagree; we lean toward keeping AI disclosure parallel to CRediT rather than inside it, with attention to the verification-and-responsibility line.

    Practical recommendations

    Three for authors. First, treat AI assistance as a tool, not a substitute. Verify, edit, and take responsibility for what appears in the paper. Second, assign CRediT roles based on what you contributed including verification, not based on what the AI produced. Third, disclose AI use in the publisher-mandated declaration; the disclosure runs parallel to CRediT, not inside it.

    Three for editors. First, treat the verification-and-responsibility threshold as the operating standard for AI-assisted contributorship. Second, require both the CRediT statement and the AI-use disclosure at submission. Third, where a contributorship statement looks like it may reflect AI-assistance role inflation, ask the standard editorial question: what did this contributor actually do?

    Three for the broader system. First, harmonise AI-disclosure formats across publishers (work the NISO and COPE community has begun). Second, maintain the contributorship-versus-disclosure separation; do not collapse them. Third, evaluate the case for a 15th CRediT role on its merits, including the costs of taxonomic expansion.

    Related dictionary entries

  • GenAI in scholarly authorship: the 2026 disclosure landscape

    The 2023 ICMJE position that generative AI cannot be a co-author has aged into a stable consensus across scholarly publishing, but the implementation surface around it has grown fast. In 2026 the question is no longer whether to disclose AI use in a manuscript; it is how, where, and with what evidence. This post maps the current disclosure landscape, the technical mitigations that publishers expect authors to apply, and the residual uncertainty around detection.

    The ICMJE 2023 position and its echoes

    In January 2023 the ICMJE updated its Recommendations to add that chatbots cannot be authors because they cannot meet the accountability criterion: an LLM cannot take responsibility for the integrity of the work, cannot approve the final version in any meaningful sense, and cannot be contacted by readers seeking clarification. The position was endorsed within weeks by the World Association of Medical Editors and by COPE. By mid-2023 every major publisher had aligned. The AI co-authorship rejection is now treated as a settled norm.

    What replaced the brief flurry of “ChatGPT as co-author” papers was a more nuanced question: how should authors disclose AI use when the system is a tool? This is where 2024 and 2025 brought significant fragmentation, and where 2026 has begun to consolidate.

    The publisher landscape in 2026

    Nature and the Springer Nature stable

    Nature requires authors to declare any use of LLMs in the Methods section (for research articles) or in the acknowledgements (for editorial and review content). The declaration must specify the model, the version, the date of use, and the purpose. Nature does not permit AI-generated images or figures except where the AI generation itself is the subject of the research. Springer Nature has cascaded a similar policy across its journals with light variation.

    Cell Press and Elsevier

    Cell Press and Elsevier journals require disclosure of AI-assisted writing in a dedicated declaration that sits alongside competing interests and funding. The declaration is structured: type of tool, purpose (e.g., language polishing, literature search, code generation, image analysis), and a confirmation that the authors take full responsibility. Elsevier additionally requires that AI-generated text be reviewed and edited by the authors and explicitly forbids using AI for peer review.

    Wiley

    Wiley’s policy distinguishes between using AI as a tool (allowed with disclosure) and using AI to generate substantive intellectual content (not allowed). The distinction is fuzzy at the boundary, and Wiley’s submission system asks authors to self-classify. Wiley also publishes its Best practice guidelines on research integrity and publishing ethics which were updated in 2024 to cover GenAI in detail.

    PLOS, eLife, F1000Research

    The open-publishing platforms have generally taken the position that AI use must be disclosed and that authors are responsible for verification, but they have been more permissive about disclosed and reviewed AI use than the closed-access incumbents. eLife in particular has experimented with AI-assisted peer review summaries, with disclosure to authors and readers.

    What “disclosure” actually requires

    The fragmentation across publisher policies has converged on a common five-element disclosure, which CASRAI’s AI disclosure helper assembles into a publisher-specific declaration:

    1. Tool and version. Not “ChatGPT” but “GPT-4o (OpenAI), version of 2025-12-04.”
    2. Purpose. One of: language polishing, translation, literature search, code generation, data extraction, image analysis, hypothesis generation, draft writing. If the use spanned multiple purposes, list each.
    3. Scope. Which sections or artefacts were involved. “Abstract and discussion polished” is meaningfully different from “first draft written.”
    4. Human verification. A statement that named authors have reviewed and verified the output and take responsibility for it.
    5. Prompt and output retention. Increasingly, journals are asking authors to retain prompts and outputs for audit. Cell Press now formally asks; Nature recommends. Treat this as a 5-year retention obligation.

    See our AI disclosure for authors guide for the publisher-by-publisher decision tree.

    The hallucination problem

    The single largest editorial concern in 2026 remains hallucination: an LLM fabricating a citation, a method, or a result and the authors failing to catch it. Retraction Watch tracked over 200 retractions in 2024 and 2025 attributable in whole or part to undisclosed AI-generated fabrications, primarily fictitious references but increasingly fabricated quantitative results in tables.

    The mitigations are well-known and surprisingly under-applied:

    • Citation verification. Every citation in an LLM-generated draft must be checked against the actual source. Tools like Scite, Semantic Scholar’s citation graph, and Crossref’s metadata API help. The bare minimum: every DOI must resolve and the paper at that DOI must say what the LLM claims.
    • Numerical verification. If an LLM produces a number, the human author must reproduce the number from the underlying source. “The LLM said it” is not provenance.
    • Retrieval-augmented generation (RAG). Grounding an LLM in a fixed corpus of verified sources, with citation chaining, reduces but does not eliminate hallucination. RAG-based research-writing tools (Elicit, Consensus, scite Assistant) have an accuracy edge over raw LLMs precisely because they constrain the model to a verifiable corpus.

    Munafò and colleagues at the UK Reproducibility Network have argued, correctly in our view, that AI-assisted writing should sit inside the same reproducibility envelope as the rest of the work: prompts and outputs are part of the methods, not part of the prose.

    Detection and watermarking

    The detection problem has not been solved. Tools that claim to identify AI-generated text by perplexity or burstiness have unacceptable false-positive rates against careful human writers and are easily defeated by simple paraphrasing. AI-assisted writing is, on the open web, essentially undetectable in 2026.

    Three more promising directions exist. First, watermarking: the major LLM providers have prototyped statistical watermarks (Google’s SynthID-Text, OpenAI’s research-stage text watermark) that embed a detectable signal in token-selection statistics without affecting fluency. Adoption has been slow because authors can defeat watermarks by re-rolling with a different model, and because no publisher has committed to refusing un-watermarked submissions. Second, provenance metadata: the C2PA standard (originally for images) is being extended to text, with cryptographically signed assertions of generation source. Third, process auditing: rather than detecting AI in the output, audit the authors’ process artefacts (version history, prompt logs, draft trail). This is the direction in which institutional integrity offices are moving.

    For authors, the practical takeaway is that you should not rely on undetectability. The conservative path is disclosure plus verification.

    What about peer review?

    The 2026 consensus is that peer reviewers may not paste unpublished manuscripts into a third-party LLM. The reason is confidentiality, not anti-AI sentiment: a paper under review is privileged information and most LLM providers retain inputs in some form. NIH and several large funders have made this an explicit policy for proposal review; publishers are catching up. eLife and a handful of others are experimenting with publisher-hosted LLM tooling that does not exfiltrate the manuscript, which threads the needle.

    Where this is going

    Three trajectories are visible. First, the disclosure form will converge: expect a NISO or COPE-led standardisation of GenAI disclosure within 18-24 months, modelled on the structured CRediT statement. Second, prompt-and-output retention will become mandatory for high-stakes journals (clinical, regulatory-relevant), and audited at random. Third, the line between “AI as tool” and “AI as substantive contributor” will be tested by hybrid systems where the human author’s contribution is curation, framing, and verification rather than generation. We expect the integrity community to draw a harder line on quantitative and methodological substance than on prose: an AI may polish your discussion section with disclosure, but an AI may not propose your analytic method without that proposal being independently validated and disclosed.

    For now, the operating rule is straightforward. If you used AI, disclose it specifically. If the AI produced text or numbers in your paper, verify them yourself. If a publisher asks for prompts and outputs, retain them. If you are reviewing a paper, do not paste it into a chatbot. The GenAI disclosure domain at CASRAI tracks the publisher-by-publisher policy text for authors who need to comply across multiple submission targets.

    Related dictionary entries

    References

    ICMJE, Recommendations (January 2023 update, defining authorship to exclude AI). WAME, Chatbots, ChatGPT, and Scholarly Manuscripts (2023). COPE, Authorship and AI tools position statement (2023, reaffirmed 2025). Nature editorial, Tools such as ChatGPT threaten transparent science; here are our ground rules for their use (2023). Munafò et al., The reproducibility debate is an opportunity, not a crisis (PLOS Biology, 2022).