Prompting, vetting, vouching. Enough for authorship?
On the use of LLMs to produce academic scholarship
I recently posted part 1 of a 2 part series on the ethics of (non-)monogamy. The second part is queued up and (almost) ready to go, but I am hereby interrupting your regular programming to share a preview of some new work on a totally different topic: the ethics of authorship in an era of large language models (LLMs).
Sorry for the whiplash!
A bit of background to get us started. My formal academic training is a mix of history and philosophy (pretty squarely in the humanities), psychology (usually seen as a social science), and cognitive science (a mutt discipline, maybe closer to the natural sciences—but all these divisions feel rather artificial).
As a consequence, I have friends, collaborators, and colleagues from most different parts of the university. And one thing I’ve noticed is that there are often different attitudes and norms around authorship, collaborative writing, and the role of large language models in scholarly publication.
Very broadly, in the humanities, the tendency is toward solo-authored pieces, with collaborative work being rarer, and LLMs seen as generally suspect (albeit, with a great deal of variance in the attitudes and approaches of individual scholars). In the social and natural sciences, by contrast, it is much more common to see multi-author papers. Intensive collaboration between people with different knowledge, training, and skill sets is normal, and even necessary for many types of projects. And LLMs, for their part, are quickly growing in popularity for all different stages of the research process.
Scholars in these fields use LLMs to brainstorm, summarise literature, test-drive arguments, structure outlines, provide stylistic suggestions, edit, and increasingly, to draft whole sections of prose in response to user prompting. These sections, in turn, can then be carefully stitched together to form a substantial portion of the first draft of an essay.
Given the current limitations of LLMs, I suspect that in most cases, these drafts will then need to undergo multiple rounds of editing “by hand” to get the material up to a publishable standard. But will that always be the case?
I’m not certain. We should, I think, consider the possibility that we are closing in on an era in which LLMs are able to produce reasonably high quality writing in response to (sufficiently sophisticated) user prompting.
If that is correct, we should not be caught flat-footed. What is going to happen when user prompted, LLM-generated text is routinely good enough to publish as it is—that is, without further writing or editing by the human user(s)? What if that “first draft” eventually doesn’t need more work to be a meaningful contribution to literature?
The status quo
Up till now, most journals and publishers have drawn a hard line around one issue: the LLM cannot itself be an author. Allegedly, this is because they are not moral agents and therefore “cannot take responsibility” for an essay’s content. Now, I am not actually convinced this judgment is correct (as Neil Levy notes, the ability to take responsibility for a text is not, in fact, a requirement of being an author in many disciplines), and I am working on a piece with some colleagues in which we hope to raise some challenges to this apparent consensus.
But suppose we accept that LLMs cannot be authors, or even co-authors—what then? There would still be an outstanding issue: What about the human user who did the prompting of the LLM? Can they be an author? Even if the text that is produced by the LLM is publishable without any further tweaking?
Can someone be an author if they didn’t write a single word?
In a new preprint with Clint Hurshman, Sebastian Porsdam Mann, and Julian Savulescu, we argue that under certain conditions, the answer is yes. If a human user makes a “substantial intellectual contribution” by conceiving of an idea, creatively prompting the LLM, vetting its output, and ultimately vouching for the material, they could in some cases count as an author according to widely accepted criteria.
(We base our argument on the criteria endorsed by the International Committee of Medical Journal Editors [ICMJE]. One could, of course, disagree with these criteria.)
Note, however, that our claim is conceptual rather than normative or prescriptive. We don’t suggest that scholars should use LLMs in this way, nor that there are no good reasons to discourage such use in various contexts. Rather, it’s that current prevailing norms of authorship in the social and natural sciences, plus medicine (among other fields) clearly allow for “authorship without writing,” in certain cases, when specific criteria are met. If we take those norms seriously, then some types of LLM-assisted work—including prompting, vetting, and vouching—should qualify for authorship too.
The “senior author” analogy
Consider a familiar situation. A principal investigator at a large medical centre outlines a thesis, provides key sources, and asks a junior colleague (perhaps a postdoc they have hired) to put together a first draft of a paper. The PI reads the draft critically, requests revisions—over iterations—ensures accuracy, and ultimately signs off on the paper. At the limit, she may never have written a sentence of the manuscript herself. Yet under ICMJE criteria she is unquestionably an author: she contributed substantially to conception and design, revised the work critically, approved the final version, and accepts accountability.
Now change the scenario only slightly. Instead of a junior colleague, the PI uses an LLM throughout an analogous (or structurally identical) process. She originates the idea, provides the structure and identifies key sources, prompts the draft, critiques it, demands revisions, verifies accuracy, and eventually signs off. If her contribution to the essay is the same as the PI who had a postdoc do the drafting, then, by parity of reasoning, she should also count as an author.
A senior author deciding whether to task an LLM or a postdoc with drafting a paper. Image created with Perplexity.
Of course, that doesn’t settle the issue. Perhaps we shouldn’t extend authorship to PIs who “merely” contribute to the conceptualisation of an idea or argument, critically review drafts, suggest changes, verify accuracy, confirm that an article moves the literature forward, and take responsibility for its content. We leave that possibility open in our essay.
And there are ethical differences between the two imagined PIs. LLMs, unlike postdocs or research assistants, cannot be mentored, for example, or otherwise helped to advance in their careers. Plausibly, replacing junior researchers with machines would erode important social goods—a concern we take seriously.
But this don’t alter the narrower point: When a human satisfies all the accepted criteria for authorship, as these are currently conceived, the presence or use of an LLM in the workflow doesn’t somehow cancel out their authorship.
Where this leaves us
Maybe this just leaves us with a dilemma. Either we accept that some forms of LLM-assisted work amount to genuine authorship by the human who prompts, vets, and vouches for the manuscript—or we revise authorship norms, potentially depriving many current researchers from being recognised as authors despite making significant scholarly contributions.
Our preprint doesn’t settle the policy questions about when LLM use should be encouraged or allowed. It simply shows that, as things stand, authorship without writing is already a feature of academic life. The consistency principle does the rest.
Preprint: Authorship Without Writing: Large Language Models and the “Senior Author” Analogy. Available on ResearchGate and SSRN.
Authors: Clint Hurshman (NUS), Sebastian Porsdam Mann (NUS; University of Copenhagen), Julian Savulescu (NUS; University of Oxford), Brian D. Earp (NUS)
Nice to see you on here, Brian! I hope your new gig in Singapore is treating you well.
LLMs force us to go back to first principles and reconsider why we care about authorship in the first place, and its purpose (either descriptively or normatively). You characterized the dilemma as one between expanding the definition to include types of LLM involvement, or restricting it at the expense of some deserving researchers. But maybe the real dilemma is whether we view authorship as a privileged category in the first place, which is directly linked to credit, status and ownership, or as something more incidental and/or arbitrary that does not in itself grant special privileges.
In my own experiences assisting an official author (including PIs), I was often struck by how subjective and entangled these divisions of labor can become, including the extent to which I felt properly credited. Part of what seems novel about LLMs in this regard is, ironically, that in replacing that interpersonal collaboration with more automated division of labor it individualizes authorship in a way that can either make "author" mean everything or nothing, depending on your perspective.
Another fruitful analogy, I think, is Supreme Court opinions. My understanding is that while they're "written" by the justices, their clerks play a role very similar to the one you're imagining for LLMs in the kind of writing process you describe. That is, the justice knows which way they want to rule and what arguments they want to use in broad strokes, but they convey that to a clerk, who writes it up, finds the relevant case law to cite, and so on. But that's just a first draft, and they'll probably have some back and forth analogous to how a human might go back and forth with an LLM.