Months before ChatGPT launched, a small group of researchers in Japan was already asking the question that would consume academic publishing for the next three years: does using AI to write research papers undermine the originality of the researcher? Nakazawa, Udagawa, and Akabayashi (2022) published their answer in mid-2022, and it’s a paper worth reading now precisely because it was written before the panic. There’s no reactionary tone, no sense of emergency. Just a careful, philosophical argument about what originality actually means in academic work and whether AI threatens it.
Their short answer is no. But the reasoning behind that answer is what makes the paper interesting, especially when you read it from the vantage point of 2026, after three years of watching institutions struggle with exactly the questions the authors raised.
AI and Researcher Originality
Nakazawa et al. (2022) break a research paper into its component parts and ask where AI support poses a genuine risk to researcher originality. Their answer is nuanced. The introduction section, which consists largely of literature review and background framing, can reasonably receive AI writing support. The researcher still reviews, modifies, and approves the text.
Even if the introduction requires some creative judgment in selecting and compiling sources, that judgment remains with the researcher. “Researchers must examine the AI-drafted text sentence by sentence, make necessary modifications, supplement citations, and approve the final version: researcher originality is thus protected” (p. 704).
The discussion section is where they acknowledge things get complicated. Nakazawa et al. (2022) recognize the discussion as the most creative part of any paper, the section where a researcher’s original thinking is most fully expressed. They identify two levels of AI involvement: checking for erroneous inferences, which poses no threat, and generating candidate interpretations from the results, which could feel like AI is doing the researcher’s intellectual work.
But even at that second level, the authors argue, the researcher still needs to evaluate, select, adjust, and approve the AI’s outputs. And if AI eventually automates even the evaluation step, a human would still need to evaluate the evaluation. Nakazawa et al. (2022) identify this as an infinite regress: at every level, someone has to exercise judgment, and that someone is the researcher.
The argument is logically clean, though it’s also somewhat abstract. In 2026, we know that the line between “evaluating AI output” and “rubber-stamping AI output” is thinner than the paper assumes, especially when researchers are under time pressure and the AI output looks good enough.
Rethinking Originality
The most provocative argument in the paper isn’t about AI at all. It’s about originality. Nakazawa et al. (2022) challenge the assumption that originality is something a researcher achieves in isolation, independent of tools and environment. They point to genome research, where studies literally cannot happen without genome analyzers. They note that researchers are constantly influenced by colleagues, conferences, readings, and everyday academic exchange. AI, in this view, becomes one more factor in a network that already shapes how researchers think and work.
The authors frame this as a move from an “individualistic” understanding of originality to a “distributed” and “collaborative” one:
The value of originality has so far tended to be “individualistic”, attributed to independent persons. Perhaps it is time to change this individualistic interpretation of originality and adopt a more “distributed” and “collaborative” meaning. Knowledge production is a collaborative process involving humans, research environments and instruments, and AI. Perhaps the dilemma of AIassisted writing of papers and diminishing authors’ originality would substantially reform the traditional knowledge framework. (p. 705)
I’ve covered two other papers that make closely related arguments from very different angles. Guetzkow, Lamont, and Mallard (2004) interviewed peer-review panelists and found that originality in the humanities and social sciences goes well beyond “new findings.” The most valued form was “original approach,” the ability to frame a problem in a new way.
Johnson-Eilola and Selber (2007) went further, arguing that “assemblage,” texts built from existing materials to solve new problems, should count as legitimate scholarly composition. All three papers, spanning two decades, converge on the same point: the lone-genius model of originality doesn’t match how knowledge is actually produced. Nakazawa et al. are just applying that insight specifically to AI.
The Authorship Standards Question
Nakazawa et al. (2022) review three major authorship frameworks: the ICMJE recommendations, Nature’s editorial policy, and the Science Council of Japan’s (SCJ) guidelines. All three require substantial intellectual contribution, drafting or critical revision, and final approval. The ICMJE and Nature add a fourth criterion: accountability for all aspects of the work, even parts the author wasn’t personally involved in.
The authors find that fourth criterion potentially too strict, since it could disqualify researchers who use AI for sections they then don’t fully understand at a technical level. This was a forward-looking concern in 2022. By 2026, the authorship question has exploded, with journals and institutions creating wildly varying policies about how AI use should be disclosed, credited, and regulated. Butson and Spronken-Smith (2024) explored exactly these implications for research practices in higher education, and the tensions they identified map directly onto what Nakazawa et al. anticipated.
The authors also suggest that AI support could actually improve scholarly work:
By obtaining AI support, researchers may be able to perform creative work in a more refined fashion. We predict that selecting AI support, evaluating it, and properly adjusting AI would remain an important aspect of work on the part of researchers. Furthermore, even if technology reaches the point where AI can do those tasks, researchers would still need to evaluate its performance. This, which leads to an infinite regress, ensures that researcher originality is protected. (p. 705)
This is a genuinely positive framing, and it was unusual in the pre-ChatGPT literature, which was already leaning toward caution and concern.
A Pre-Panic Paper, Read After the Panic
Nakazawa et al.’s paper was published in September 2022, two months before ChatGPT launched. The AI tools they were discussing were earlier, less capable systems, not the generative models that would soon reshape every conversation about academic writing. And yet, the core argument holds up remarkably well.
What hasn’t held up is the assumption that researchers will always carefully evaluate AI output before using it. The infinite regress argument, someone always needs to judge the judge, is logically sound but practically optimistic. We’ve seen enough cases of researchers submitting AI-generated text with fabricated references, placeholder language, and even ChatGPT’s own self-identification phrases to know that the “careful evaluation” step gets skipped more often than the authors anticipated. The originality argument works, but only if researchers actually do the work of evaluation. When they don’t, what you get isn’t distributed originality. It’s outsourced production.
Still, the paper’s central contribution holds: originality was never the product of a single mind working in isolation. Researchers have always relied on tools, environments, collaborators, and institutions. AI is a new kind of tool, a powerful one, but the question was never whether it belongs in the process. The question is how researchers use it, and whether they maintain the intellectual engagement that makes the work genuinely theirs.
The authors were asking the right question before most people knew it needed asking. That counts for something, even if the answer needs updating now that we’ve seen what happens when the tools get powerful and the standards stay vague.
References
- Butson, R., & Spronken-Smith, R. (2024). AI and its implications for research in higher education: A critical dialogue. Higher Education Research & Development, 43(3), 563-577. https://doi.org/10.1080/07294360.2023.2280200
- Guetzkow, J., Lamont, M., & Mallard, G. (2004). What is originality in the humanities and the social sciences? American Sociological Review, 69(2), 190–212. http://www.jstor.org/stable/3593084
- Johnson-Eilola, J., & Selber, S. A. (2007). Plagiarism, originality, assemblage. Computers and Composition, 24(4), 375–403. https://doi.org/10.1016/j.compcom.2007.08.003
- Nakazawa, E., Udagawa, M., & Akabayashi, A. (2022). Does the use of AI to create academic research papers undermine researcher originality? AI, 3(3), 702–706. https://doi.org/10.3390/ai3030040