Authorship Without Writing: Large Language Models and the Senior Author Analogy

Brian D. Earp; Clint Hurshman; Julian Savulescu; Sebastian Porsdam Mann

REVIEW 3 major objections 4 minor 18 references

Authorship Without Writing: Large Language Models and the Senior Author Analogy

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A researcher can be the author of an LLM-written paper without writing a word, provided she guides the work, reviews it critically, and takes responsibility for it.

desk verdict Useful reframing of the LLM-authorship debate, but the key move—that the drafting entity's nature is irrelevant—is stipulated, not defended, and the paper's own vetting concession blurs the analogy. read the letter →

arxiv 2509.05390 v1 pith:ID5SIRFH submitted 2025-09-05 cs.CY cs.AIcs.CL

Clint Hurshman , Sebastian Porsdam Mann , Julian Savulescu , Brian D. Earp This is my paper · ORCID

classification cs.CYcs.AIcs.CL

keywords researchethicsauthorshipcriterialargelanguagemodelsseniorLLM-assistedwritingICMJEpublicationphilosophicalargument

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Many published papers list senior researchers as authors even though they never wrote any of the text; what matters is guiding the work, reviewing it critically, approving the final version, and taking responsibility. This paper argues that a researcher who does the same things with a large language model—prompting it, critiquing its draft, and approving the final text—is an author in exactly the same sense. The argument runs by analogy: if the senior author Smith is an author, and Jones's contribution to her LLM-assisted paper is identical in every respect that feeds into the usual authorship criteria, then Jones is an author too. The upshot is that LLM use, even to generate a complete draft, can be legitimate authorship under current norms; the only consistent alternative is to change those norms, which would strip authorship from many existing senior authors.

What carries the argument

The argument turns on paired case studies—'Senior Author' Smith and 'LLM user' Jones—constructed so that the two researchers give identical instructions, receive the same draft, and give the same critical feedback. The load-bearing identity is the ICMJE four-criterion test, used as a filter that decides which differences between cases are 'relevant' to authorship: conception, critical revision, final approval, and accountability. Because the only difference between the cases is the nature of the assistant (human postdoc vs. LLM), and because that difference is not among the authorship-relevant categories, the two cases must receive the same verdict.

What would settle it

A systematic survey of journal authorship policies in bioethics and biomedical research could settle premise 1: if a substantial share of journals require authors to have directly drafted at least some text, then Smith fails the very criteria the analogy rests on, and Jones falls with her.

Watch

Extended reading notes

Core claim

The paper's central claim is that the prevailing, widely accepted authorship criteria (exemplified by the ICMJE four-part test) make authorship a function of conception, critical revision, final approval, and accountability—not of producing words. Since a paradigmatic senior author meets those criteria without writing, and since a researcher who guides an LLM can meet the same criteria in the same way, consistency requires recognizing the LLM user as an author. The paper formalizes this as a modus ponens: Smith is an author; if Smith is an author then Jones is an author; therefore Jones is an author. It also draws the contrapositive: if Jones is not an author, current authorship norms need f

Load-bearing premise

The argument holds only if Smith and Jones are identical in every way that matters for authorship, and the paper defines 'what matters' by the ICMJE categories—so the difference between a human assistant and a non-agent LLM is treated as irrelevant without independent proof.

Editorial extensions

If this is right

A researcher can be listed as (sole) author of a paper they never wrote a word of, as long as they conceived the project, critically reviewed LLM drafts, approved the final version, and accepted accountability.
If that conclusion is rejected, the rejection cannot stop with LLM users: current authorship criteria in medicine, bioethics, and many sciences would also stop covering senior authors who delegate drafting to a junior collaborator.
Irresponsible LLM use—submitting output without critical review and without accepting accountability—fails the authorship criteria, just as a senior author who rubber-stamps a postdoc's draft would fail them.
The paper's conclusion does not settle whether using LLMs is wise, fair, or ethical; those questions turn on separate empirical and normative considerations.
Formal role disclosure (such as CRediT statements) and AI-use statements are compatible with the argument and can keep credit honest while the human remains the author.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the analogy is right, the same logic should extend to other non-human drafting tools—automated analysis scripts, grammar engines, or future AI—so authorship attribution would track guidance and oversight rather than the mechanism of text production.
The argument predicts a testable pattern in empirical authorship-attribution studies: perceived authoriality should track conception, critical revision, and accountability more than word-level composition, so readers should rate Jones similarly to Smith when both roles are disclosed.
The paper leaves open a practical design question: journals may want a CRediT-style role that names 'conceptual guidance and critical revision of AI-generated text,' so that the human's authorial contribution is visible without requiring the LLM to be an author.
A deeper tension: if LLM output demands more vetting than a postdoc's draft, then Jones may need to do more authorial work than Smith to reach the same confidence; the paper treats this as making Jones 'more authorial,' but a stricter reading would make the two cases only similar in degree, not identical.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The thing to know about this paper: it's a clearly argued conceptual piece that makes a genuine contribution by drawing out the senior-author analogy. The authors separate three questions (can humans be authors, can LLMs, should they), and focus on the first. They build a transparent modus ponens: Smith the senior author is an author without writing a word; Jones the LLM user contributes the same way; therefore Jones is an author. The 'minimal senior authorship' framing is new and useful, and the objection-and-reply section is thorough—especially the treatment of mentorship and social goods, which they correctly keep separate from authorial status.

The soft spots are real but not disqualifying. The load-bearing premise is the relevance filter: that the nature of the drafting assistant (human postdoc vs. non-agent LLM) is irrelevant to the human's authorship. The paper asserts this rather than argues for it. The stress-test is right that Section IV.3 concedes Jones must vet more critically than Smith to responsibly take accountability, which is a difference in the 'revising critically' clause. The reply that Jones is thereby 'more authorial' shifts from exact analogy to graded similarity without a stated threshold. That said, the idealized cases are stipulated to be identical, and the concession only says real-world Jones has more work to do; it doesn't obviously destroy the analogy, but it does reveal that the relevance filter is doing all the work. Also, footnote 4 asserts that ICMJE criteria are sufficient for authorship, which is a contested normative assumption; the paper flags it but doesn't defend it beyond pragmatic reasons. A missing engagement with the updated ICMJE guidance on AI tools (which postdates the 2019 version they cite) is a minor gap.

The disjunctive conclusion ('either Jones is an author or authorship norms need revision') is weaker and still depends on the same relevance filter, so it doesn't independently rescue the modus ponens. But the paper's value isn't only in the proof; it's in the reframing, and the authors are honest about the normative assumptions.

Who's it for: anyone working on research ethics, authorship policy, or AI ethics. It deserves serious peer review—not a desk reject. A good referee would push on the relevance filter and ask for a principled criterion. I'd bring it to a reading group.

Referee Report

3 major / 4 minor

Summary. The paper argues that a human researcher who uses an LLM to generate a complete research manuscript can nonetheless be an author, by analogy with a senior author who does not write any words. It introduces two parallel cases (Smith supervising a postdoc; Jones supervising an LLM), argues that Smith meets the ICMJE authorship criteria, claims that Smith and Jones contribute to their respective papers in the same relevant respects, and concludes by modus ponens that Jones is an author. The paper then replies to objections: denying premise 1, the disanalogy of social goods, the responsibility/hallucination problem, and the lack of moral agency in LLMs. It ends with a disjunctive conclusion: either LLM-assisted authorship is legitimate or current authorship norms require fundamental revision.

Significance. If the argument were successful, it would have direct implications for publication ethics, credit allocation, and the concept of authorship in biomedical and scientific research. The paper is clearly structured, takes major objections seriously, and is unusually candid about the limits of its own argument, including the empirical literature on LLM hallucination and the context-dependence of authorship norms. It also contains a transparent AI-use declaration. The central philosophical claim is timely and important, and the case-study method makes the argument accessible. However, the inference rests on at least two load-bearing assumptions that are asserted rather than adequately defended: the sufficiency of the ICMJE criteria and the identity of Smith's and Jones's contributions in the relevant respects. These need repair before the argument can be considered sound.

major comments (3)

[Section III.2 and Section IV.3] The justification of Premise 2 is undermined by the paper's own concession that Jones must engage more critically with LLM output than Smith must with a postdoc's draft. Section II stipulates that Smith and Jones provide the same parameters, receive the same draft, and provide the same feedback, but Section IV.3 says that 'in real cases, this may not strictly be true' and that Jones likely faces a higher burden due to LLM hallucination. If Jones must do more vetting to responsibly take responsibility, then her contribution is not identical in the relevant respect, and the exact analogy collapses. The reply that this makes Jones 'more authorial' (IV.3) shifts from a binary identity claim to a graded comparison without supplying a threshold. Consequently, the disjunctive conclusion ('either Jones is an author or current criteria require revision') is not independently secured: if Premise 2
[Footnote 4 and Section III.1] The argument treats the four ICMJE criteria as sufficient for authorship, but this sufficiency is asserted rather than defended. The footnote offers two pragmatic reasons—mitigating unfair denial of credit and preventing strategic omission—but these do not establish conceptual sufficiency. If the criteria are only necessary, then Smith's satisfying them does not entail that Smith is an author, and Premise 1 fails. The paper does not engage with alternative authorship standards, such as a requirement of direct textual contribution or intentional authorship, except to dismiss them as contextually inapplicable. Since the entire modus ponens rests on this step, the sufficiency claim needs a more robust defense.
[Section IV.3] The treatment of the responsibility objection is ambiguous about whether the accountability criterion is necessary for authorship. The paper cites Levy (2025) to contest the ICMJE accountability requirement, but then continues to rely on Jones's responsibility-taking as part of what makes her 'more authorial.' If accountability is not required, the extra vetting burden is irrelevant to authorship; if it is required, then the case description must explicitly specify that Jones in fact performs the additional vetting beyond what Smith does. As written, the paper leaves the status of Premise 2 unclear and does not provide a determinate answer.

minor comments (4)

[Section III.2] The sentence 'though we consider possible disanalogies in Section III.2' appears to be a cross-reference error; the disanalogies are considered in Section IV.2 (and IV.3).
[Section IV.4] The word 'artesanal' should be 'artisanal.'
[General] The paper uses 'senior author' to characterize Jones, but Jones is a sole author in the LLM case. The terminological relationship between 'senior authorship' and 'sole authorship' should be clarified early, since the analogy to Smith's co-authorship is not exact on its face.
[Section IV.4] The discussion of externalist reasons to consider LLMs as full-fledged authors is presented as 'somewhat weighty' but is then set aside as beyond scope. This digression could be trimmed or more clearly separated from the central argument, as it risks confusing the paper's main claim about the human user's authorship.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central analogy argument is grounded in an external authorship benchmark and does not reduce to its own conclusion by definition or self-citation.

full rationale

The paper’s core inference is a modus ponens over an external benchmark: Smith satisfies the ICMJE authorship criteria (Box 1, Section III.1), Jones is stipulated to make the same relevant contributions (Section III.2), so Jones satisfies the same criteria. The controversial step is not a hidden reuse of the conclusion but the asserted analogy between Smith’s and Jones’s contributions. An asserted premise may be contestable, but it is not circular. Section IV.3 concedes that Jones may need to vet LLM output more carefully than Smith vets a postdoc’s draft; the paper treats this as making Jones more authorial rather than as collapsing the argument into its conclusion. That is a weakness in the analogy’s strength, not a circularity. Self-citations (e.g., Porsdam Mann et al. 2023a, 2023b, 2024; Earp et al. 2024) appear in background literature or in supporting asides, and none is load-bearing for the central modus ponens. The paper also explicitly frames its conclusion as conditional on accepting existing ICMJE-style norms and offers the modus tollens alternative that the norms should be revised, so it does not define the desired conclusion into existence. No fitted parameters, no uniqueness theorem, and no definitional equivalence between input and output are present. The main risk is philosophical—whether the relevance filter is adequately defended—not circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper's claim rests on normative and interpretive premises about what counts as authorship, not on empirical fits or new entities. There are no free parameters and no invented entities. The load-bearing assumptions are the sufficiency of the ICMJE criteria, the interpretation of 'substantial contribution,' the irrelevance of the drafting entity's nature, and the exclusion of mentorship from authorial relevance. All are named in the text and all are contestable; the paper flags the first two explicitly.

assumptions (4)

domain assumption The four ICMJE criteria are sufficient as well as necessary for authorship.
Footnote 4 states 'we take it that the ICMJE's guidelines are intended as sufficient conditions as well as necessary ones.' The modus ponens needs sufficiency; if the criteria are only necessary, Smith's being an author does not establish Jones's.
domain assumption Providing a thesis, rough structure, and representative citations counts as a 'substantial contribution' to conception or design under ICMJE criterion 1.
Box 1 and Section III.1 apply criterion 1 in this way; the paper offers no independent account of what makes a contribution substantial, and a critic could deny that such thin input qualifies.
domain assumption The drafting assistant's nature (a human moral agent vs. a non-agent LLM) is irrelevant to whether the human is an author, as long as the human performs the relevant authorial acts.
This is the core of premise 2 (Section III.2, 'identical in all relevant respects') and of the reply in Section IV.4 that the human's intention 'imbues' the LLM output. If the assistant's agency or accountability is deemed relevant, premise 2 fails.
domain assumption Social goods such as mentorship and collaboration are not part of the authorial contribution and therefore cannot disanalogize Smith and Jones.
Section IV.2 classifies mentorship as a non-authorial difference; the paper argues this, but the relevance of authorship to the exclusion of social goods is a normative commitment without an independent derivation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Authorship Without Writing: Large Language Models and the Senior Author Analogy." pith.science (2026). https://pith.science/paper/ID5SIRFH

@misc{pith2026250905390,
  author       = {Pith},
  title        = {Pith review of: Authorship Without Writing: Large Language Models and the Senior Author Analogy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ID5SIRFH}},
  note         = {Machine review of arXiv:2509.05390}
}

read the original abstract

The use of large language models (LLMs) in bioethical, scientific, and medical writing remains controversial. While there is broad agreement in some circles that LLMs cannot count as authors, there is no consensus about whether and how humans using LLMs can count as authors. In many fields, authorship is distributed among large teams of researchers, some of whom, including paradigmatic senior authors who guide and determine the scope of a project and ultimately vouch for its integrity, may not write a single word. In this paper, we argue that LLM use (under specific conditions) is analogous to a form of senior authorship. On this view, the use of LLMs, even to generate complete drafts of research papers, can be considered a legitimate form of authorship according to the accepted criteria in many fields. We conclude that either such use should be recognized as legitimate, or current criteria for authorship require fundamental revision. AI use declaration: GPT-5 was used to help format Box 1. AI was not used for any other part of the preparation or writing of this manuscript.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[1]

senior author

AI was not used for any other part of the preparation or writing of this manuscript. Keywords: Research ethics, professional ethics, philosophy, biomedical research\ This is a preprint of a paper that has been submitted to a journal. It has not yet gone through peer review. It may be cited as follows: Hurshman, C., Porsdam Mann, S., Savulescu, J., & Earp,...

work page 2025
[3]

How Smith meets the ICMJE criteria for authorship The ICMJE requires that an author meet all four of the following criteria (2019 version):

work page 2019
[4]

The norms surrounding research produced with the help of generative artificial intelligence (AI) are less clear in comparison (Ganjavi et al

Criterion 4 (Accountability): ✔ By attaching her name, she accepts responsibility for the integrity of the work. The norms surrounding research produced with the help of generative artificial intelligence (AI) are less clear in comparison (Ganjavi et al. 2024, An et al. 2025). Since the release of ChatGPT in November 2022, researchers have incorporated it...

work page 2024
[5]

2024, Khan et al

In what ways, if any, should human researchers use LLMs to write research papers? Regarding the third question, the use of LLMs raises a variety of ethical concerns, not only about distribution of credit (Earp et al. 2024, Khan et al

work page 2024
[7]

[c]ontributors who meet fewer than all 4 of the above criteria for authorship should not be listed as authors, but they should be acknowledged

provide some crucial guidance. Specifically, the ICMJE recommends that authorship be based on four criteria, all four of which must be met by each individual who can legitimately be counted as an author. As a reminder, these are: ● Substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data fo...

work page 2014
[8]

substantial contribution

is not completely value-neutral, insofar as it takes existing authorship norms, as exemplified by the ICMJE guidelines, for granted. Authorship norms are contested, so one could fairly argue that these guidelines should be revised, but this, too, is a separate issue. Our conclusion is therefore partly an ameliorative one: that Jones has the kind of author...

work page 2021
[9]

authorship,

IV.2 Denying Premise 2 Alternatively, one could deny premise 2 by insisting that there is a disanalogy between the contributions of Smith and Jones to their respective research papers. Most obviously, while Jones’s authorship is fundamentally a solitary process (i.e., on the assumption that LLMs are not sentient agents and do not provide the same kind of ...

work page 2024
[10]

senior author

will lead not only to the atrophying of users’ own ability to write, but also their ability to critically engage with LLM-produced texts to ensure their integrity (see Voinea et al. in preparation). Our tendency to depend on cognitive scaffolds when doing so is reasonably effective can easily lead to overreliance, though this may be unintended and even un...

work page 2025

Show all 18 references

[13]

2025, August

https://dailynous.com/2021/10/14/co-authorship-in-philosophy-over-the-past-120-years-bourget-weinberg/ Brynjolfsson, E., Chandar, B., & Chen, R. 2025, August

2021
[14]

https://www.sustainability.google/reports/google-2025-environmental-report/ Grice, H

2025 Environmental Report. https://www.sustainability.google/reports/google-2025-environmental-report/ Grice, H. P

2025
[16]

Paper under review

Your brain on ChatGPT: Accumulation of cognitive debt when Preprint. Paper under review. using an AI assistant for essay writing task. Preprint available at arXiv: https://arxiv.org/abs/2506.08872 Lang, B. H., Nyholm, S., & Blumenthal-Barby, J

arXiv
[17]

Preprint available at arXiv: https://arxiv.org/abs/2411.05025 Ostertag, G

LLMs as Research Tools: A Large Scale Survey of Researchers' Usage and Perceptions. Preprint available at arXiv: https://arxiv.org/abs/2411.05025 Ostertag, G

arXiv
[18]

AI & SOCIETY

Bullshit universities: the future of automated education. AI & SOCIETY. https://doi.org/10.1007/s00146-025-02340-8 Stokel-Walker, C. 2023, January

2023 doi
[1975]

grasping

comparable to that of a human coauthor. By analogy, some of the authors of this paper have argued that LLMs can play the role of identifying relevant reasons in an argument which can 6 Because they lack interests, LLMs are also incapable of benefiting from credit for authorshi...

2024
[2006]

Bioethics 20(4): 213-220

Author, contributor or just a signer? A quantitative analysis of authorship trends in the field of bioethics. Bioethics 20(4): 213-220. Bourget, D., & Weinberg, J. 2021, October

2021
[2014]

conceptual

found that empirical papers in bioethics had 2.97 authors on average, while “conceptual” papers had 1.35. In philosophy—as in most other humanities disciplines—coauthorship was historically uncommon, due both to institutional pressures and a less-clear division of labor than i...

2021
[2024]

AI & Ethics

Engaging the many-hands problem of generative-AI outputs: a framework for attributing credit. AI & Ethics. https://doi.org/10.1007/s43681-024-00440-7 Klarna. 2024, February

2024 doi
[2025]

writing is thinking

but also about potential deskilling or even wholesale replacement of human labor by generative AI in certain domains, which is Preprint. Paper under review. increasingly occurring in creative, customer-service, and programming work (Klarna 2024, February 27; AbuMusab 2024, Bry...

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

senior author

AI was not used for any other part of the preparation or writing of this manuscript. Keywords: Research ethics, professional ethics, philosophy, biomedical research\ This is a preprint of a paper that has been submitted to a journal. It has not yet gone through peer review. It may be cited as follows: Hurshman, C., Porsdam Mann, S., Savulescu, J., & Earp,...

work page 2025

[2] [3]

How Smith meets the ICMJE criteria for authorship The ICMJE requires that an author meet all four of the following criteria (2019 version):

work page 2019

[3] [4]

The norms surrounding research produced with the help of generative artificial intelligence (AI) are less clear in comparison (Ganjavi et al

Criterion 4 (Accountability): ✔ By attaching her name, she accepts responsibility for the integrity of the work. The norms surrounding research produced with the help of generative artificial intelligence (AI) are less clear in comparison (Ganjavi et al. 2024, An et al. 2025). Since the release of ChatGPT in November 2022, researchers have incorporated it...

work page 2024

[4] [5]

2024, Khan et al

In what ways, if any, should human researchers use LLMs to write research papers? Regarding the third question, the use of LLMs raises a variety of ethical concerns, not only about distribution of credit (Earp et al. 2024, Khan et al

work page 2024

[5] [7]

[c]ontributors who meet fewer than all 4 of the above criteria for authorship should not be listed as authors, but they should be acknowledged

provide some crucial guidance. Specifically, the ICMJE recommends that authorship be based on four criteria, all four of which must be met by each individual who can legitimately be counted as an author. As a reminder, these are: ● Substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data fo...

work page 2014

[6] [8]

substantial contribution

is not completely value-neutral, insofar as it takes existing authorship norms, as exemplified by the ICMJE guidelines, for granted. Authorship norms are contested, so one could fairly argue that these guidelines should be revised, but this, too, is a separate issue. Our conclusion is therefore partly an ameliorative one: that Jones has the kind of author...

work page 2021

[7] [9]

authorship,

IV.2 Denying Premise 2 Alternatively, one could deny premise 2 by insisting that there is a disanalogy between the contributions of Smith and Jones to their respective research papers. Most obviously, while Jones’s authorship is fundamentally a solitary process (i.e., on the assumption that LLMs are not sentient agents and do not provide the same kind of ...

work page 2024

[8] [10]

senior author

will lead not only to the atrophying of users’ own ability to write, but also their ability to critically engage with LLM-produced texts to ensure their integrity (see Voinea et al. in preparation). Our tendency to depend on cognitive scaffolds when doing so is reasonably effective can easily lead to overreliance, though this may be unintended and even un...

work page 2025