The ethics and potential of writing academic papers with generative AI

Introduction

Professor Atsushi Mizumoto of the Faculty of Foreign Language Studies at Kansai University has released a free guide titled "Ethical and Efficient Academic Writing in English Using Generative AI" (https://langtech.jp/ai-writing/). Based on a 2025 English seminar hosted by the Kansai Branch of the Information Processing Society of Japan, the guide systematically compiles practical prompt examples tailored to the IMRaD structure along with ethical precautions.

Within a day of its release, the X post announcing it received over 767 likes and more than 79,000 views, drawing attention from many researchers and graduate students.

Here I outline the guide and draw on related X posts and public information to examine the challenges of using generative AI ethically.

Structure and key points of the Mizumoto guide

The guide consists of five sections.

1. Introduction, ethics, and policy — Current status and potential of generative AI, academic integrity and plagiarism prevention, guidelines from publishers and educational institutions

2. Basics of English paper structure and corpora — IMRaD structural patterns, checking formulaic expressions with corpora, style, tense, logical flow, and AI prompts

3. Generating drafts through dialogue with AI — Prompts for refining English text, writing strategies for each section (Introduction, Methods, Results, Discussion)

4. Quality control and submission preparation — Grammar checks and style consistency, fact-checking and reference formatting, declaring AI use

5. Summary and practical guide — Common failure patterns, checklists, workflows and list of tools

Particularly important is the concept of "Augmented Competence." This approach treats AI as a tool to extend human capabilities, limits its use to idea generation and drafting, and places final responsibility on the person using the AI. Rather than leaving the writing to AI, the most important part of ethical use is for the writer to think for themselves while working with AI.

What the NIH restriction on grant applications shows

The restriction introduced by the National Institutes of Health (NIH) in the United States is an important case to consider when looking at generative AI use in research. After a single principal investigator (PI) used generative AI to prepare and submit more than 40 grant applications, the NIH moved to restrict the submission of applications created with generative AI.

This case is a classic example of easily giving in to the temptation to mass-produce applications using generative AI. I take seriously the fact that the same problem is occurring in paper writing. If low-quality papers apparently written with generative AI are submitted in large numbers, the burden on editors and peer reviewers will likely increase substantially.
Peer review has traditionally been carried out by many researchers on a volunteer basis. If submissions of low-quality AI-generated papers increase, the peer review system itself could become unsustainable.

AI hallucinations found in submitted papers

This concern is already materializing. According to an investigation of papers submitted to ICLR 2026 (reported in December 2025), AI hallucinations (such as fabricated citations) were detected in 50 out of a sample of 300 papers (16.7%), including papers that had received high review scores.

These errors were overlooked even after review by three to five experts. Because AI-generated papers read plausibly, even human reviewers struggle to spot mistakes.
Automation of peer review itself using AI is also advancing, but we are still far from a stage where summarization and novelty checks by LLMs can be standardized. For now, human eyes remain indispensable.

Ethics at the stage of conceiving research

The Mizumoto guide is valuable for systematically presenting the ethics of how to write. When we consider the big picture of research activities, however, what we need next is ethics "before writing." I hope we will also see texts in the future that delve into how to use generative AI ethically when conceiving research worth publishing.

Limitations of RLHF (Reinforcement Learning from Human Feedback) in LLMs

LLMs are trained to prioritize "plausible" answers and tend to correct novel ideas as "errors" (the False-Correction Loop). This tendency points to the problem of leaving the entire task of coming up with ideas to AI.

The problem of mass-producing "shallow ideas"

Mass-producing ideas with AI leads to more research plans that look well-structured on the surface but lack depth. Grant reviewers and journal editors must evaluate and review all of them, which increases their workload.

How to use AI ethically

The principle of "Augmented Competence" also applies to generating ideas. Rather than having AI come up with ideas, you present your own ideas and ask for feedback. Asking AI, "What are the weaknesses of this idea?" or "How much does this overlap with existing research?" is useful, whereas delegating the whole task by asking AI to "give me 10 research ideas" risks lowering the quality of research.

Impact on peer reviewers and editors

Academic publishing is supported by researchers who volunteer to do peer review. This system is fundamentally based on goodwill and trust, and the mass submission of low-quality AI-generated papers shakes that trust to its core.
On X, academic journal editors have reported seeing more plausible-sounding research derived from generative AI in peer review. For reviewers, spending time reading papers that look the part but are empty inside is a heavy burden.

Hopes for the future: Texts that address "ethics before writing"

Professor Mizumoto's guide is valuable in showing how to use AI ethically and efficiently in writing papers.
While building on it, we also need to deepen ethical discussions on how to use AI to conceive research ideas before writing a paper. Key issues include the following:

• Using AI appropriately as a "sounding board" when generating research ideas — Using AI to test and deepen your own ideas rather than having it create them

• Guarding against the risk of mass production — Practicing self-restraint: just because you can easily produce more does not mean you should

• Responsibility to the academic ecosystem — Taking care that your research activities do not become a burden on reviewers and editors

• Disclosing AI use — Making a habit of stating clearly what AI was used for and at which stage.

Conclusion

Generative AI is a powerful tool for researchers. Misused, however, that power could erode the foundations of scholarship.
Professor Mizumoto's guide showed how to use AI ethically and efficiently in writing papers. Taking its lessons on board, we also need to deepen discussions on ethics "before writing"—that is, the ethical use of AI in creating research ideas.
I believe this is necessary to protect the health of scholarship, to protect reviewers and editors, and to protect the value of our own research.

Related

What healthcare professionals should know first when using generative AIshoei05.com

Using generative AI to prepare grant applicationsshoei05.com