How should graduate students use AI so as not to tread the line: a practical guide to maintaining academic integrity

The well-known academic integrity tracking website Retraction Watch published an in-depth report titled “The case of the fake references in an ethics journal (opens in a new tab)” on December 2, 2025. “Journal of Academic Ethics (opens in a new tab)” (Journal of Academic Ethics) published by the internationally renowned publisher Springer Nature was found to have at least 19 fabricated references out of 29, and 18 of the Google Scholar links were all empty. The dark humor lies in this: the problem was not that the author did not understand ethics, but that he sent the AI-generated quotes directly, without doing one thing less: verifying them one by one.

You may feel that this is far away from you. However, in November of the same year, a demography paper co-authored by a scholar and a doctoral student at the University of Hong Kong was revealed to have extensively cited ghost documents (opens in a new tab) invented by AI. The University of Hong Kong subsequently intervened in the investigation and announced the punishment results in December. It is also reported that some scholars in Taiwan discovered that a master’s thesis quoted a work that he had never written. None of these people want to fake, they just outsource the judgment that should be made by themselves to the model.

Have you ever counted how many citations in your last literature review were found for you by AI but you did not click on the original text to confirm?

What this article wants to tell you is a boundary

I’m not going to talk about how to use AI to write faster. What I want to talk about is another thing: to what extent AI can be used in every aspect of the paper, so as to amplify your productivity and maintain academic integrity. After reading, you will have a clear boundary and a set of division of labor principles that you can follow.

To put it bluntly, the real issue of using AI for graduate students has never been about whether it can be faster, but about which links can be used and to what extent?

Before taking action: write down the tripartite consensus first

Before opening any AI tool, keep the bottom line of academia in mind. This is not a rule of one academic journal, but a common premise of the entire field.

COPE (Committee on Publication Ethics) (opens in a new tab) made it clear in 2023: AI tools cannot satisfy authorship because it cannot take responsibility for submissions and cannot handle conflicts of interest and authorship. The general policies of ICMJE, Elsevier, and Nature are consistent: AI cannot be named as an author, its use must be disclosed, and researchers are ultimately responsible for all content. Taiwan is no exception. Ministry of Education’s Academic Ethics Education Resource Center (opens in a new tab) proposes five major considerations: verification, copyright, personal information protection, avoidance of over-reliance, and self-disclosure; the National Science Council’s usage guidelines state that according to its policy requirements, AI output cannot be fully trusted, and the final judgment cannot replace human independent thinking.

Condensing this entire paragraph into one principle that can be stuck to the side of the screen, that is: AI is a magnifying glass, not a substitute. Verifiability and accountability remain with you forever.

Five links, where are the boundaries drawn?

Next, let us look at it link by link to see where this line should be drawn?

Step one: Literature exploration, use it to find follow-up readings, do not outsource judgment

Let AI help you condense a stack of documents into summaries, sort out the disputes between different schools of thought, and expand keywords you haven’t thought of. These are all very useful. However, there is a red line here: any reference list that AI directly spits out to you should be treated as pending verification and not citation-ready.

The previous number is not to scare you. Early research found that about 55% of the references in the literature review generated by ChatGPT-3.5 were fabricated, and even if it was switched to GPT-4, nearly 20% would still be wrong. For every citation, you must go back to the original database and click on the original text to confirm that it really exists and really says what you think it says.

Step 2: Research the design and treat it as a dialectical partner, not a substitute.

This step is best for letting AI act as your most critical reviewer. Throw it your research question and ask it to ask you: Why is this question important? Are your variables endogeneous? Are there alternative explanations that you haven’t thought of?

The value of helping you find blind spots is far greater than helping you create a design. You have to make your own decision on the design itself. After all, you are the one who defends every choice at the oral interview table, not the model.

Step 3: Data analysis, assistance is possible, fabrication is not possible

Let AI help you organize the verbatim draft into preliminary coding, translate the statistical results into vernacular, and check for inconsistencies in coding before and after. These are all reasonable accelerations. But there is one rule NEVER: don’t let it embellish or complete your information. It has no idea what your data looks like, it just produces something that looks reasonable. Once these results are included in the paper, it will be your name.

Step 4: Writing and disclosure, distinguish between polishing and ghostwriting

Where is the line between language polishing and ghostwriting? My judgment is this: the ideas, arguments, and structure are yours, and the AI ​​only cares about how to explain them clearly. This is called embellishment; if it even generates the arguments, it has crossed this line.

After polishing, remember to reveal. Exposing is not a confession, it is making it clear what you used and where you used it, so that others can reproduce and test it. The requirements of most academic institutions or academic journals can be condensed into one sentence: you must be able to explain every part of the research that benefits from AI, and there cannot be paragraphs that even you can’t explain clearly.

The reason why I want to talk about this part in particular is because it is the question I have been asked most frequently since the beginning of the course. In the AI Academic Writing Course (opens in a new tab), almost every graduate student asked me the same thing: Is it OK to use AI to help me polish, especially the English paragraph? Or will it be counted as ghostwriting as soon as AI intervenes? What everyone is really stuck on is never whether to use it or not, but where to draw the line at this level. My answer still goes back to the previous sentence: if it touches on expression, it is still on the polishing side; if it touches on ideas and arguments, it is inappropriate and you must go through it yourself.

Step 5: Submission and oral examination, let it be your sparring partner

This is the most underrated use of AI. Before submitting your paper, ask it to simulate the reviewers of the target journal and write a rejection letter for your paper. It will list out the most likely points first, so that you will have time to make up for them. Before the interview, let him play the most difficult oral defense, asking you three rounds of questions, forcing you to stress test your argument.

This usage does not hurt integrity at all, because in the end it is you who stands on the stage and answers. Please remember: AI is just a training partner, not a test taker.

The three most common questions

**If AI is used, should we talk about it? Will it appear that I am not independent enough? **

Think about it backwards. If you hide it and don’t tell it, it will be fatal if you are discovered. Research estimates that in 2024, at least about 10% of papers will have traces of large-scale language model processing, but they are rarely exposed. This is why various companies have tightened their policies. Proactively and clearly exposing it is evidence that your research is rigorous enough.

**Opponents say that AI will hurt research and training. Is this true? **

It’s true, you have to face it honestly. The study “Your Brain on ChatGPT (opens in a new tab)” published in June 2025 by Nataliya Kosmyna, a scientist at the Media Laboratory of the Massachusetts Institute of Technology (MIT), explored the brain wave response of humans when using large language models (LLM). Research has found that people who use large-scale language models have the weakest connections in the brain areas, have the lowest sense of ownership of their own work, and even have difficulty quoting accurately what they have just written. Researchers call it “cognitive debt.”

But this is not a reason not to use it, but a reason to use it. What these findings really say is: Outsourcing the most important aspects of judgment, criticism, or argumentation is equivalent to short-circuiting the development of your research abilities. So the line is drawn here: leave it to yourself if it’s worth using your own brain; only when it comes to running errands, searching or translating, can you consider leaving it to AI.

**Different journals have different regulations, how can I keep up? **

Just remember the minimum consensus: AI cannot be anonymous, it must be exposed, and people are fully responsible. The details vary, with some requiring prior approval from the instructor and others requiring only subsequent disclosure. Spending ten minutes reading the original AI policy of the target journal before submitting is far more practical than memorizing all the rules.

Next step: Do this little exercise today

If you only want to take away one thing from this article, then take away this sentence: The highest standard for making good use of AI is that every paragraph of content can be explained, every document can be found, and the person with the name can afford it.

Open the manuscript you are writing, find all the paragraphs that AI has generated or rewritten for you, and ask yourself those three questions one by one. If there is something you can’t answer, it means you haven’t really turned it into your own research.

Please leave the part where you need to use your own brain the most. This part requires you to do it yourself.

---

If you want to turn this set of boundaries into a checklist that you can check before submitting your paper, I have compiled an “AI Academic Integrity Self-Checklist” to help you make sure that every citation can be found, every paragraph is clearly explained, and the person named can afford it, paragraph by paragraph, before submitting your paper. Welcome to click here to obtain it for free and incorporate this boundary into your research routine.