b19 — Fact-sheet 3 — Historical authorship-paradigm precedents#

Compiled:

2026m05d12

Compiled by:

Claude Opus 4.7 Max (subagent for the b19 AI co-authorship analysis)

Scope:

Bourbaki, software tools, AlphaFold/early AI tools, megacollaborations, deceased authors

Methodology:

Primary sources retrieved via WebFetch/WebSearch where possible; URLs cited per claim

Status:

Independent reference document — descriptive, not prescriptive

Reader’s note

This fact-sheet describes five precedents in scientific authorship that bear structural analogy (sometimes positive, sometimes negative) to the AI-co-author case. It does NOT recommend any path for any specific paper.

A note on retrieval method

WebFetch (direct page retrieval) was unavailable for this compilation, so primary-source claims below are drawn from WebSearch result snippets that quote the underlying primary pages (ICMJE, COPE, ATLAS, CMS, LIGO, IPCC, Nature, Wolfram support, AMS biographical articles). Verbatim quotation that could not be reconstructed from snippets is flagged with [QUOTE NEEDS VERIFICATION]. All claims below should be checked against the cited URLs before quotation in any externally facing document.

(a) Nicolas Bourbaki — the collective pseudonym#

Historical and structural facts#

  • Founding and identity. “Nicolas Bourbaki” is a collective pseudonym adopted in mid-1930s France by a group of young mathematicians, most of them alumni of the École Normale Supérieure in Paris. The would-be members met for the first time in late 1934 in a Parisian café to plan a rigorous treatise on analysis; the name was taken from a French general of the Franco-Prussian War (1870–71) via an earlier student prank in which the senior student Raoul Husson delivered a parodic lecture attributing fake theorems to “General Bourbaki” and similar names. The first name Nicolas was supplied by Eveline de Possel, wife of founding member René de Possel, who thus became the pseudonym’s “godmother.” Sources: Wikipedia Nicolas Bourbaki; MacTutor Bourbaki 1; Quanta Magazine, Inside the Secret Math Society Known Simply as Nicolas Bourbaki (2020-11-09); Britannica.

  • Founding members. The core founders included Claude Chevalley, André Weil, Henri Cartan, Jean Dieudonné, and several others (sources give “eight or nine”). Membership was kept deliberately fluid; new members were elected by the existing group. Sources: Wikipedia; CNRS News, Bourbaki and the Foundations of Modern Mathematics.

  • Mandatory retirement at 50. A retirement rule “at or about 50 years of age” was reportedly introduced in 1953 and enforced from 1956, justified by Weil’s claim that mathematicians are most creative in their twenties and thirties. The aim was to force “gradual disappearance” of the founding members and require generational renewal. The historian Liliane Beaulieu has noted she found no written affirmation of the rule and that exceptions occurred — so the rule is best read as a strong norm rather than a constitutional clause. [FACT NEEDS VERIFICATION] for the exact wording of the rule. Sources: Wikipedia; encyclopedia.com The Bourbaki School of Mathematics.

  • Treatment as a real person by editors. Bourbaki applied for individual membership in the American Mathematical Society twice (1948 and 1950). The AMS Secretary J. R. Kline, who already knew Bourbaki was a collective, rejected the application on the ground that Bourbaki “was not an individual” and invited the group to reapply at the institutional membership rate. Bourbaki later declined. Sources: TCNJ Department of History, Taking Nicolas Bourbaki Personally (2016-02-08); Pieronkiewicz, Mathematical Communities: Mathematicians Who Never Were.

  • Individual contributors named? No. The defining feature of the Bourbaki convention is that the published treatises (Éléments de mathématique and the Séminaire Bourbaki lectures) bear only the collective pseudonym on the title page; current membership has been kept secret on principle. Individual contributors are identified retrospectively by historians, by personal disclosure, or via the Bourbaki archive (the “less secret Bourbaki archive” referred to by Lieven Le Bruyn), but not in the publication record itself. Sources: Wikipedia; neverendingbooks, “The (somewhat less) Secret Bourbaki Archive.”

Motivating principles#

As stated by participants and reconstructed by historians:

  1. Mathematical unity / depersonalisation of authority. The treatise was meant to present mathematics as a single structured edifice from an axiomatic viewpoint, not as a collection of personal achievements. A single pseudonymous “author” enforces stylistic uniformity and removes per-author authority claims.

  2. Cross-generational accumulation. The retirement-at-50 rule and the recruitment of younger members were designed to allow the “author” Bourbaki to persist while no individual human did. The work was longer than any career.

  3. Insulation from personality cult. Hiding membership prevented the work from being read as the opinion of any particular school or person, and (per Mashaal and others) protected internal debate from external careerist pressure.

  4. Playful resistance to bureaucratic categories. The AMS correspondence shows Bourbaki cheerfully exploited the gap between “individual” and “institution” rather than resolving it.

Structural analogy to AI co-authorship#

The analogy holds on these axes:

  • Both Bourbaki and a named AI co-author would be a non-individual signing entity on the title page.

  • Both confront editors with the problem that standard individual-accountability mechanisms (one person, one signature, one ORCID) do not map cleanly onto the entity in question.

  • Both raise the question of whether the output can carry an authorship credit that no single biological person could carry alone.

The analogy breaks on these axes:

  • Bourbaki is a covering name for known humans who could in principle be unmasked and held accountable. The accountability is hidden by convention but exists in fact. An AI system is not a covering name for any human and has no analogous “behind-the-curtain” individual whose membership in a profession or institution could be invoked.

  • Bourbaki had no incentive structure that depended on credit attribution: members had separate careers under their own names. An AI system has no career, no reputation that survives, and no external coverage that depends on the credit.

  • Bourbaki’s pseudonym was deliberately stable across decades; the identity persists by recruitment and retirement. A given AI model version is a snapshot — it does not “retire” in the Bourbaki sense, it is replaced, and the replacement is not a peer-elected continuation.

  • The AMS rejection itself (Bourbaki “is not an individual”) is the same structural objection that ICMJE, Nature, Science and others raise against AI authorship; the Bourbaki precedent shows that the category objection has operated as a binding constraint for at least seventy-five years, not just since 2023.

(b) Software tools — citation but not authorship#

Historical and structural facts#

  • Standard practice. Statistical packages (R and its CRAN packages, SAS, SPSS, Stata, Python’s scientific stack — NumPy, SciPy, pandas, statsmodels, scikit-learn) and symbolic-computation tools (Mathematica, Maple, SageMath) are cited as software, not listed as authors. The norm is enforced by ICMJE-compatible authorship policy at major journals.

  • Wolfram’s official citation guidance. Wolfram Research instructs users to cite Mathematica “just as you would reference a book or any other publication,” in the form Wolfram Research, Inc., Mathematica, Version [number], Champaign, IL ([year]). Source: Wolfram Support Quick Answers #472, How do I reference Wolfram products in papers?

  • R citation infrastructure. R provides the built-in citation() and citation("packagename") functions explicitly to produce a BibTeX entry for use in papers. The R Core Team and individual package maintainers appear as authors of the citation for the software, not as co-authors of the downstream paper. Sources: rOpenSci, How to Cite R and R Packages (2021-11-16); CRAN report package vignette Report and Cite Packages.

  • Was the tool ever proposed as an author and rejected? The pre-LLM software-tool literature does not record serious proposals to list a statistical package itself as a named author on a research paper. The norm of citing-rather-than-co-authoring software appears to have been settled without controversy until generative AI raised the question in 2023.

  • AI-tool variant (since 2023). When ChatGPT was listed as a co-author on at least four 2022–2023 preprints and papers, Nature and Science moved within weeks to prohibit AI tools as listed authors. Nature: “no LLM tool will be accepted as a credited author on a research paper, because any attribution of authorship carries with it accountability for the work, and AI tools cannot take such responsibility.” Science: “AI-assisted technologies … do not meet the journals’ criteria for authorship and therefore may not be listed as authors or co-authors.” ICMJE updated its guidance in 2023 to the same effect, requiring instead disclosure in the Methods and/or Acknowledgements. Sources: Nature, Tools such as ChatGPT threaten transparent science; here are our ground rules for their use (2023); Science / AAAS blog, Change to policy on the use of generative AI and large language models (2023); ICMJE, Recommendations (current).

Motivating principles#

As stated by participants and editors:

  1. Accountability. ICMJE criterion 4 requires the author to be “accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.” Software (and AI systems) cannot answer post-publication queries, sign retractions, or assume legal/professional liability.

  2. Approval of final version (criterion 3). Software cannot give final approval of the version submitted in any sense that survives editorial scrutiny.

  3. Citation suffices for credit and reproducibility. Software citation gives the version number, the vendor or maintainer, and a retrievable reference. This is what is structurally needed for replication. Authorship credit is for the people who designed the study and judged its results.

  4. Conflict-of-interest disclosure. An author can disclose financial or institutional conflicts; a software package cannot.

Structural analogy to AI co-authorship#

The analogy holds on these axes:

  • An AI assistant used during a research project is, structurally, software invoked at runtime. The default category fit is “cite-as-software.”

  • ICMJE criterion 4 (accountability), criterion 3 (final approval), and the conflict-of-interest mechanism all map cleanly: each is a binding constraint that LLMs cannot satisfy as autonomous agents.

  • The reproducibility argument transfers: what downstream readers need is which model version, with what prompt, in what session, not a name on the byline.

The analogy breaks on these axes:

  • Mathematica is not asked to make scientific judgements; the judgement of which transformation to apply is the user’s. LLM-grade assistants can be deployed in judgement-shaped roles: drafting argument structure, raising objections, classifying evidence. Whether this difference of kind matters is exactly the question the precedent does not settle, because it never arose for Mathematica.

  • A software package has a static behaviour; an LLM session is conversational and adaptive. The “tool” framing assumes the user directs the tool. In long sessions the influence can run bi-directionally — which is what tempts some authors to consider listing the model as a contributor.

  • The structural reply remains: even if the role is judgement-shaped, the accountability mechanism still requires a human signatory. This is the line that Nature, Science and ICMJE have drawn.

(c) AI tools and AlphaFold — credit pattern in published work#

Historical and structural facts#

  • The AlphaFold paper itself. Jumper, J., Evans, R., Pritzel, A., et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). DOI: 10.1038/s41586-021-03819-2. Published online 15 July 2021. URL: https://www.nature.com/articles/s41586-021-03819-2. The paper has approximately 32 named human authors, all DeepMind-affiliated, with John Jumper and Demis Hassabis in lead positions (Jumper first, Hassabis last). The methods section runs roughly 60 pages and describes the 32-component algorithm. The acknowledgements thank a long list of named DeepMind colleagues and project managers, plus the JAX, TensorFlow and XLA teams. AlphaFold itself is the subject of the paper, not an author of the paper. Sources: Nature article record; Wikipedia AlphaFold; cross- referenced WebSearch snippets of the acknowledgements section.

  • Predecessor paper. Senior, A.W. et al. “Improved protein structure prediction using potentials from deep learning.” Nature 577, 706–710 (2020). Same pattern: AlphaFold v1 is the object described, not a listed author.

  • Downstream use papers. Papers that use AlphaFold to predict structures cite the Jumper 2021 paper (and the AlphaFold Protein Structure Database, hosted at EBI: https://alphafold.ebi.ac.uk/). EBI maintains an explicit “How to cite AlphaFold” page that lists the Jumper et al. 2021 and Varadi et al. 2022 references. AlphaFold is not listed as an author of any downstream paper that this compilation could identify. [FACT NEEDS VERIFICATION] — no exhaustive search was performed for counter-examples. Source: EBI training page How to cite AlphaFold.

  • DeepMind as an author? DeepMind is the affiliation of the human authors of Jumper 2021. It is not listed as a corporate author on the byline. The “AlphaFold” name does not appear as an author on either the source paper or, so far as identifiable, on downstream user papers.

  • AI tools credited as more than software — documented cases. The strongest documented public cases of AI being treated as more than software are the brief 2022–2023 episodes in which ChatGPT was listed as a co-author on a small number of preprints and early papers. Major publishers (Nature, Science, JAMA, Elsevier, Springer Nature, the ICMJE) repudiated this convention within months, replacing it with mandatory disclosure in Methods or Acknowledgements. Sources: Nature, ChatGPT listed as author on research papers: many scientists disapprove (2023-01); NIEHS Environmental Factor, March 2023.

  • Schoenfeld 1985 precedent. Schoenfeld, A. H., Mathematical Problem Solving. Academic Press, Orlando, FL (1985). ISBN 0-12-628870-4. This book is best known for its protocol analysis of human problem-solving behaviour and its four-component framework (resources, heuristics, control, beliefs). The WebSearch corpus available here did not surface a specific section of Schoenfeld 1985 in which a calculus tutoring computer is credited as a co-author or in any other non-tool role. The book’s conceptual contribution to AI-assisted teaching is via the framework that intelligent tutoring systems subsequently adopted, not via a co-authorship convention for the tutor itself. [FACT NEEDS VERIFICATION] — the specific Schoenfeld passage referred to in the prompt could not be located through search snippets alone; a direct page reference from the book would be needed to characterise it accurately. Sources: Schoenfeld 1985 (Academic Press); Schoenfeld 2016 reprint in NASSP Bulletin; instructionaldesign.org Mathematical Problem Solving (A. Schoenfeld).

Motivating principles#

As stated by publishers and the AlphaFold team:

  1. Object vs. author. When the AI system is the subject of investigation, it is described in the paper, not credited as an author of the paper. This is the same convention as for any instrument, algorithm or organism under study.

  2. Tool use disclosure. When the AI system is used as a tool during research, current Nature/Science/ICMJE policy requires disclosure of the system, version, and use, in Methods or Acknowledgements — but explicitly not in the author list.

  3. Accountability barrier (recurring). The same accountability argument from precedent (b) applies: an AI system cannot satisfy ICMJE criterion 4.

  4. Corporate co-authorship not used as a workaround. DeepMind / Google DeepMind is not listed as a corporate author of the AlphaFold paper. The convention is human researchers, affiliated with DeepMind. This stands in contrast to (d) below, where corporate / collaboration authorship is the operating norm.

Structural analogy to AI co-authorship#

The analogy holds on these axes:

  • The AlphaFold 2021 paper is the canonical recent example of an AI system being credited for its scientific contribution without being listed as an author. The contribution is documented in the paper’s content and method, and the credit flows to the human team via the byline.

  • The ChatGPT-as-author episode is the canonical recent example of the opposite convention being attempted and rejected by major publishers within months.

The analogy breaks on these axes:

  • AlphaFold 2021 is a paper about the AI system itself. For papers in which an AI co-thinker contributed to the scientific argument rather than being the artefact under study, AlphaFold is not the precedent — the closer (and less favourable) precedent is the rejected ChatGPT-as-author episode.

  • The Schoenfeld 1985 reference, as the prompt poses it, would be a useful precedent if the book did treat a computer tutor as a contributor; this could not be confirmed from search snippets and needs a direct page reference before being used.

(d) Megacollaboration author lists — ATLAS, CMS, LIGO, IPCC#

Historical and structural facts#

CERN ATLAS#

  • Size. Approximately 5,500 members and approximately 3,000 scientific authors as of recent rosters. Source: ATLAS public pages, https://atlas.cern/.

  • Authorship convention. “All ATLAS CONF and PUB notes must have as authors ‘ATLAS Collaboration’, without explicit names.” [QUOTE NEEDS VERIFICATION] — paraphrased from ATL-GEN-PUB-2008-001 (the ATLAS authorship policy document on cds.cern.ch). For published papers, the named author list is generated automatically from the membership roster on a “reference date” (the date of the first circulation of the draft to the collaboration); the list is “frozen” at the second circulation and used for final submission. Sources: ATL-GEN-PUB-2008-001, https://cds.cern.ch/record/1110290/; ATLAS public Authors pages.

CERN CMS#

  • Authorship convention. The published author list is “alphabetical by country, then alphabetical by institute, then alphabetical by author name.” Authors begin signing one year after joining the collaboration and stop signing one year after leaving. An Authorship Committee maintains the list. Sources: CMS public pages; CMS TWiki EprRulesExplained; cernopendata GitHub issue #642 CMS author lists: order of authors?.

LIGO Scientific Collaboration (LSC)#

  • Authorship convention. The author list is alphabetical and includes engineers and technicians who contributed in important ways to the design, construction, installation, commissioning, or operation of the detectors and major LSC facilities. For some classes of papers, “Group 1” authors are listed first in an order chosen by themselves, followed by “Group 2” authors alphabetically. Source: LIGO-T010168, LSC Publication and Presentation Policy, https://dcc.ligo.org/public/0026/T010168/.

  • Compilation. Author lists are versioned and published as separate LIGO documents (e.g. LIGO-M1300559, LIGO-M1600027), with cut-off dates for who appears on a given paper based on membership status and effort fraction.

IPCC#

  • Author roles. The IPCC explicitly distinguishes:

    • Coordinating Lead Authors (CLAs) — responsible for coordinating a chapter;

    • Lead Authors (LAs) — responsible for production of designated sections;

    • Contributing Authors (CAs) — provide text, graphs or data for assimilation by the LAs;

    • Review Editors (REs) — ensure that comments are appropriately handled;

    • Chapter Scientists — technical and logistical support.

    CLAs and LAs together “have collective responsibility for the contents of a chapter.” Source: IPCC, Role of CLA, LA, RE document, https://www.ipcc.ch/site/assets/uploads/2017/08/Role_of_CLA_LA_RE.pdf; IPCC, How does the IPCC select its authors? factsheet (2021); https://www.ipcc.ch/about/preparingreports/.

PubMed / MEDLINE indexing#

  • MEDLINE indexes a group (corporate) author name alongside personal authors when present in the byline. From May 2006 onward, corporate authors are displayed in the order in which they appear in the published byline. When a group author name appears, individual group members listed in the article are indexed as collaborators (PubMed search tag [ir]), not as authors. Sources: NLM, Authorship in MEDLINE; NLM Technical Bulletin 2008 Mar–Apr; NLM Technical Bulletin 2006 May–Jun.

  • Bibliometric studies report that group-authored articles in PubMed grew from approximately 1.04% (2000–2004) to approximately 1.35% (2015–2019), and that the number of papers with more than 1,000 authors has more than doubled in the past five years. Sources: Nature (2019), Hyperauthorship: global projects spark surge in thousand-author papers; Cronin’s foundational work on hyperauthorship.

What “accountability for all aspects of the work” means in this regime#

ICMJE criterion 4 — “Agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved” — has been the focus of explicit critique in the hyperauthorship literature. Several adaptations have emerged:

  1. Distributed accountability. A named author is accountable for the portion of the work they certify (their analysis, their detector subsystem, their chapter section), and accountable for the collaboration’s process for handling questions about other parts.

  2. Collaboration-level accountability. ATLAS publications are signed “ATLAS Collaboration” for notes, with explicit named authors for journal papers. The collaboration as a whole carries process-level accountability through its review committees.

  3. Author-list governance. Authorship Committees (CMS), Publication Committees (LIGO), and review editor structures (IPCC) provide a named institutional locus for handling integrity questions — replacing what an individual author would otherwise do.

  4. Documented criticism. Bibliometric studies have argued that no individual on a 3,000-author particle-physics paper can substantively satisfy criterion 4 in the strict sense, and that the criterion is reinterpreted in practice rather than literally satisfied.

Motivating principles#

  1. No individual could be substantively responsible for the full work. A multi-billion-dollar detector or a multi-thousand-page assessment is by construction a collective product. The author list is a roster of contributors rather than a list of co- judging principal investigators.

  2. Credit must follow effort. Engineers, technicians, and long-term operational contributors are named because the instrument they built is what produced the result.

  3. Collaboration governance replaces individual approval. Where ICMJE criterion 3 (final approval) cannot be carried by each author individually, an internal review process (CMS analysis review, ATLAS editorial board, LIGO Publication and Presentation Committee, IPCC government and expert review) takes on the function.

  4. Searchability and credit-attribution problems are accepted as the cost of doing the work. Bibliometric studies (cited above) document that individual contributions on hyperauthored papers are hard to attribute; the collaborations have accepted this in exchange for the science.

Structural analogy to AI co-authorship#

The analogy holds on these axes:

  • Megacollaborations have already broken the ICMJE-4 assumption that every author can personally vouch for the entire paper. There is therefore a working precedent for redistributing accountability away from the literal text of criterion 4 onto a governance structure that handles integrity questions.

  • Collaboration-as-author conventions (“ATLAS Collaboration”, “LIGO Scientific Collaboration”, “IPCC Working Group I authors”) show that the byline can carry a non-individual signing entity when there is an institution behind it.

  • Distinct author roles (CLA / LA / CA / RE / Chapter Scientist) show that scientific publishing already handles role-differentiated contribution.

The analogy breaks on these axes:

  • The collaboration-as-author convention works because the collaboration is staffed by accountable humans with institutional affiliations. An AI co-author has no institutional staffing behind it of that form. (The closest analogue would be naming Anthropic — but Anthropic is not the AI, it is the manufacturer, closer to “DeepMind affiliating the human authors of AlphaFold” than to “ATLAS Collaboration” — see (c).)

  • Engineers and technicians on a LIGO author list are humans whose contribution is to the physical instrument and to its operation. An LLM is neither an instrument operator nor a one-time builder of the apparatus; the analogy to engineering credit does not slot cleanly.

  • The role-differentiation precedent (IPCC’s CA vs LA vs CLA) might in principle support a “Contributing Author” role for an AI, but current ICMJE guidance explicitly forbids this for the accountability reasons in (b).

(e) Deceased-author rules#

Historical and structural facts#

  • The ICMJE position. ICMJE’s four authorship criteria do not mention posthumous inclusion. Strictly applied, criterion 3 (“Final approval of the version to be published”) and criterion 4 (“Agreement to be accountable for all aspects of the work…”) cannot be satisfied by a deceased author for a posthumously-finalised paper. The ICMJE Recommendations therefore do not directly authorise posthumous authorship; the practice exists in the gap. Source: ICMJE Recommendations, Defining the Role of Authors and Contributors, https://www.icmje.org/recommendations/.

  • The COPE position. COPE has published a case discussion, Author deceased prior to submission, https://publicationethics.org/guidance/case/author-deceased-prior-submission, in which the consensus advice is that the co-authors who know the deceased author’s contribution best are the right judges of whether to retain the name; the journal should not over-ride that judgement from outside; and a footnote should record the deceased status and date of death. [QUOTE NEEDS VERIFICATION] for the exact COPE wording. Source: COPE case guidance (URL above).

  • Operating convention at journals.

    • Most journals (Science Editor, BMJ, ASM, JMIR, IOP and others) advise retaining the deceased author’s name if the author met the substantive contribution criteria (criteria 1 and 2) before death.

    • A footnote on the title page identifies the author as deceased, typically with a dagger † against the name and a footnote giving the date of death.

    • Consent for posthumous authorship is handled by the surviving co-authors and, where practical, the next-of-kin or estate. Some journals require a signed statement from the corresponding author certifying that the deceased author had reviewed an earlier draft and would have approved the version submitted, or a statement that the surviving authors take responsibility for final approval on the deceased author’s behalf.

    • Bibliometric studies show that publications by deceased authors are increasing in volume; the operational rules have stabilised even though the ICMJE base text is silent.

    Sources: Science Editor, The Authorship of Deceased Scientists and Their Posthumous Responsibilities; BMJ 2024, Ethics of posthumous scholarly authorship in the sciences; PLOS ONE 2022, Perish and publish: Dynamics of biomedical publications by deceased authors; ASM Journals authorship policy.

  • Structural argument for the exception. The deceased-author carve-out is defensible on the following structural logic:

    1. The deceased author made a real, identifiable contribution that meets criteria 1 and 2 before death.

    2. Criterion 3 (final approval) is satisfied either by prior approval of an earlier draft (and good-faith stewardship by the survivors) or by a named co-author’s certification that the changes since the deceased author’s last review are editorial rather than substantive.

    3. Criterion 4 (accountability) is redistributed: the surviving co-authors carry the post-publication accountability burden for the whole work, including the deceased author’s portion.

    4. A visible marker (dagger + footnote) preserves transparency for the reader.

    The deceased-author rule is thus, formally, an example of redistribution of accountability from criterion 4 onto a smaller surviving group, justified by criteria 1 and 2 having been met during the contributor’s lifetime.

Motivating principles#

  1. Credit must follow real contribution. Refusing to name a contributor merely because they died first would punish them for timing.

  2. Accountability remains binding but transferable. Criterion 4 is not abandoned; it is borne by the surviving authors on the deceased author’s behalf.

  3. Visibility, not erasure. The dagger-and-footnote convention ensures that readers and downstream auditors know which author was deceased and when.

  4. Consent is reconstructed from the historical record. Prior approval of an earlier draft is treated as evidence of consent to the substantive content; survivors handle the final-version judgement.

Structural analogy to AI co-authorship#

The analogy holds on these axes:

  • This is the single ICMJE-recognised case in which an entity that cannot personally satisfy criteria 3 and 4 at the moment of publication is nonetheless retained on the byline.

  • The mechanism that makes it work — transferred accountability to named living co-authors, with a visible marker of the non-standard status — is structurally available for other non-standard signatories.

  • Both deceased authors and AI systems fail criterion 4 in the literal sense of being able to respond to a post-publication integrity query.

The analogy breaks on these axes:

  • A deceased author was an accountable individual at the time of the substantive contribution. ICMJE criteria 1, 2 and (often) 3 were satisfied while the contributor was alive and competent. An AI system was never an accountable individual in the same legal / professional sense.

  • Survivors of a deceased author can speak for the deceased author from prior knowledge: they know what the deceased would have said. An AI model’s “prior intent” is harder to characterise because the model has no stable selfhood across sessions and no career to defend.

  • Consent for posthumous authorship has a documented historical record (the prior draft, the co-author’s recollections). Consent by an AI model has no analogous record.

  • The visible-marker convention (dagger + footnote) could in principle be borrowed for an AI contributor (e.g. an explicit named indication that an entity on the byline is an AI), but no major journal currently allows this; ICMJE, Nature and Science forbid AI on the byline outright as of 2023–2024.

Cross-case summary#

The five cases are summarised below in a single table for cross-case comparison. The “structural analogy suggests…” column is descriptive of where the analogy points, not a recommendation for any specific paper.

Five precedents in scientific authorship#

Case

Solution adopted

Governing principle

The structural analogy to AI co-authorship suggests…

  1. Bourbaki pseudonym

Collective pseudonym on the title page; members kept secret; mandatory retirement at 50; individual names never appear on Éléments de mathématique.

Mathematical unity and depersonalisation; cross-generational accumulation; the work outlasts any individual; resistance to personality cults.

A non-individual signing entity can in principle persist on a byline across generations. But the AMS rejection (1948, 1950) shows that journals/societies have been treating “is the entity an individual?” as a binding categorial question for three-quarters of a century already.

  1. Software tools

Cite as software in the references; disclose use; do not list on the byline.

ICMJE criterion 4 (accountability) and criterion 3 (final approval) cannot be met by software. Citation supplies version-level reproducibility.

The default category fit for an AI assistant is “cite as software / disclose in methods.” The 2023 publisher reaction to ChatGPT-as-author re-affirmed this category fit.

  1. AlphaFold and early AI use

AI is the subject of the paper, not an author; human researchers (DeepMind staff) appear on the byline; AlphaFold is cited in downstream user papers.

The “object vs author” distinction; institutional affiliation does the work of corporate credit; accountability flows through the human team.

Strong precedent for crediting AI content without naming the AI as author. The Schoenfeld 1985 thread, as posed, was not confirmable from available search snippets and needs direct page-reference verification.

  1. Megacollaborations

Collaboration-as-author byline (ATLAS Collaboration, LIGO Scientific Collaboration); alphabetised personal name lists with cut-off dates; role-differentiated authorship in IPCC (CLA / LA / CA / RE / Chapter Scientist); MEDLINE corporate author indexing.

Some scientific work cannot be carried by a single accountable individual. Governance structures (Authorship Committees, Publication Committees, Review Editor systems) carry the accountability that ICMJE-4 places on the individual.

Redistribution of accountability from an individual to a governance structure is a working pattern in current scientific publishing. But the governance structure in every megacollaboration is staffed by humans with institutional affiliations; there is no precedent for staffing it with the signing AI itself.

  1. Deceased authors

Retain the deceased author’s name on the byline if criteria 1–2 (substantive contribution) were met during life; dagger footnote with date of death; surviving co-authors take on criterion-4 accountability.

Credit must follow real contribution; accountability is binding but transferable to survivors; visible marker preserves transparency.

This is the single ICMJE-tolerated case of an entity remaining on the byline despite being unable to personally satisfy criteria 3–4 at publication. The mechanism — transferred accountability + visible marker — is structurally portable. The strongest disanalogy is that the deceased author was an accountable individual at the time of contribution; an AI never was, and consent cannot be reconstructed from a historical record.

References#

Primary policy documents#

Megacollaborations#

AlphaFold#

Bourbaki#

Software citation and AI tool policy#

Deceased authors and hyperauthorship#

Schoenfeld#

  • Schoenfeld, A. H., Mathematical Problem Solving. Academic Press, Orlando, FL (1985). ISBN 0-12-628870-4.

  • Schoenfeld, A. H., Learning to Think Mathematically: Problem Solving, Metacognition, and Sense Making in Mathematics (reprint 2016). NASSP Bulletin.

On this page
Recent Posts
Tags