TLDR: A new preprint argues that human researchers using large language models (LLMs) to generate papers should be considered authors, drawing an analogy to traditional “senior authors” who guide junior researchers without writing the text themselves. The paper asserts that if the human user provides the core ideas, critically reviews the output, and takes full responsibility, their contribution aligns with established authorship criteria, suggesting that either LLM-assisted authorship is legitimate or current norms require significant revision.
The use of large language models (LLMs) in academic writing has sparked considerable debate, particularly concerning who should be credited as an author. A recent preprint paper, “Authorship Without Writing: Large Language Models and the “Senior Author” Analogy”, by Clint Hurshman, Sebastian Porsdam Mann, Julian Savulescu, and Brian D. Earp, delves into this controversy by drawing a compelling parallel between LLM users and traditional “senior authors” in research.
The core of the paper’s argument rests on an analogy: just as a senior author can guide a junior researcher to produce a paper without writing a single word themselves, a human researcher using an LLM can achieve authorship under similar conditions. The authors highlight that in many scientific and medical fields, authorship is already distributed among large teams, where some contributors, including senior authors, primarily conceptualize, supervise, and vouch for the integrity of the work rather than drafting the text.
To illustrate this, the paper presents two scenarios. In the “Junior Author” case, Professor Smith, a Principal Investigator, instructs her postdoc, Charlie, to write a draft based on her ideas and structure. Smith then critically reviews, suggests improvements, and ultimately approves the final manuscript, taking responsibility for its content. She is considered a senior author despite not writing the words herself.
In the analogous “LLM” case, Professor Jones, also a PI, instructs an LLM (named “Charlie”) to generate a draft using her parameters. Jones then critically evaluates the LLM’s output, provides feedback for revisions, and approves the final version, taking sole responsibility. The paper argues that if Smith is considered an author, then Jones, whose contributions are functionally equivalent in terms of ideation, supervision, critical engagement, and accountability, should also be recognized as an author.
The authors refer to the widely accepted criteria for authorship set by the International Committee of Medical Journal Editors (ICMJE), which requires four conditions: substantial contribution to conception/design or data, drafting or critical revision for intellectual content, final approval, and accountability for all aspects of the work. They demonstrate how both Professor Smith and Professor Jones meet all these criteria, even without directly writing the text.
The paper also addresses several common objections to LLM-assisted authorship. One objection concerns the lack of direct writing, to which the authors respond by reiterating that “authorship” in research is a technical notion that accommodates diverse contributions beyond mere text composition. Another concern is the social disanalogy, where LLM use might diminish mentorship opportunities. While acknowledging this, the authors argue that such ethical considerations, while important, do not negate the authorial status itself. They also point out potential benefits, such as enabling researchers in under-resourced institutions to produce papers.
Regarding the responsibility for potential “hallucinations” or inaccuracies by LLMs, the paper stresses that the human user (like Jones) has a heightened responsibility to critically vet the LLM’s output. This critical review is, in fact, a key component of meeting authorship criteria. Finally, the objection about LLMs lacking moral agency or sentience is addressed by clarifying that the paper focuses on the human user’s authorial status, not the LLM’s. The human user, by intentionally prompting, reviewing, and endorsing the output, imbues the text with their authorial intention.
Also Read:
- Unmasking ‘LLM Hacking’: The Hidden Threat to Scientific Conclusions from AI Annotations
- Unmasking the Flaws: A Deep Dive into AI Scientist Systems’ Hidden Pitfalls
In conclusion, the paper suggests that consistency with existing authorship norms, particularly those recognizing senior authorship, implies that LLM-assisted writing, when responsibly managed, constitutes a legitimate form of authorship. The authors propose that either this form of authorship should be recognized, or current authorship criteria need fundamental revision, which could impact many human researchers currently considered authors.
