Using AI in academic writing and its ethical boundaries

A graduate student working with books and a laptop in a university library

When I talk with graduate students, the topic of AI now comes up almost every time. Some say they have never used these tools and are afraid they would get into trouble if they did. Others let on that they use them at every stage, from the literature review to interpreting their analyses, but have never told anyone. Neither extreme is healthy. Here I want to share a few basic points about using AI in academic writing, especially for graduate students working on a thesis or an article.

First, let’s be clear that using AI in academic writing is not a single thing. There is a big difference between having a tool fix the grammar of a paragraph and having it write an entire chapter. Chatting with a tool to understand a concept is also not the same as putting the literature summary it produced into your thesis. When we discuss ethics, we need to distinguish the type and degree of use.

In general, the uses most likely to be accepted are those that improve the form of content you produced yourself. Checking grammar and spelling in a text you wrote, getting suggestions to make a long sentence more readable, or improving the fluency of an abstract you wrote in a foreign language fall into this group. The ideas, arguments and interpretations are yours. The tool only polishes the wording.

Even with something as harmless-looking as language editing, I recommend being careful. While “improving” a sentence, these tools can replace an established term in your field with an everyday word, turn a carefully hedged statement into a firm claim, or erase a fine distinction between two concepts. Reading and accepting each suggested change one by one is much safer than having the whole text rewritten at once. Texts that pass through the same tool also start to sound alike, and developing your own academic voice is part of graduate education.

Then there are uses that support learning. Asking a tool to explain a statistical concept you find difficult from different angles, discussing the strengths and weaknesses of a research design, or thinking through objections that could be raised against your argument can speed up your learning when done well. But here too you need to verify what the tool says against primary sources. It does not replace a methods textbook.

The problematic zone is where the tool starts thinking and producing in your place. Having a tool write your research question, theoretical framework, interpretation of findings or discussion section and presenting it as your own text with minor edits violates the basic assumption of academic writing: that the person whose name is on the text owns and is responsible for those ideas. The real point of a thesis is for you to grow as a researcher. When you hand that process over to a tool, you may end up with a thesis, but not with the skills that thesis was supposed to give you.

Fabricated references and verification

One of the most dangerous aspects of AI in academic writing is that it can produce sources that do not exist. These tools can suggest entirely fabricated references with realistic-looking author names, journal titles, years and even page numbers. Sometimes they pair a real author’s name with an article that does not exist, and sometimes they attribute a finding to a real article that the article does not contain.

The second case is more insidious than the first. Because the source really exists, it may pass your checklist, yet the claim you base on it in your text may not appear in it at all. When a committee member or reviewer opens that source, the situation that emerges is a serious academic integrity problem, even if it was not intentional.

So my rule is very simple: do not cite a source you have not read. If an AI tool suggested a source, find it in library databases, open it and read the relevant part yourself. Make sure the source actually supports your claim. Do your literature review through systematic searches in your field’s databases, not with these tools. At most, the tools can play a supporting role, such as suggesting search terms.

The same care applies to analyses and findings. When you ask a tool to interpret your statistical output, its answer may sound very reasonable and still recommend the wrong test, misread a value or mention a relationship that is not in your data. You need to be able to explain every step of your analyses. At a thesis defence, the answer to “why did you choose this test?” cannot be “the tool said so.”

Disclosure and data privacy

Stating your use openly sits at the centre of the ethical discussion. Whether using AI tools in academic writing is acceptable depends largely on transparency. Declaring which tool you used, for what purpose and at what stage gives both readers and reviewers the chance to evaluate your text properly.

Institutions and journals differ on this, and the rules are changing quickly. So rather than memorising a general rule, I suggest checking the relevant policy every time. Before submitting an article to a journal, read that journal’s policy on AI use. If you are writing a thesis, look at the guidelines of your graduate school and university. If you cannot find a rule, talk to your supervisor and declare your use anyway. A hidden use that comes to light later can have far heavier consequences than the same use openly declared.

Listing an AI tool as an author, on the other hand, is not an acceptable approach. Authorship means taking responsibility for the work, being accountable for its mistakes and being able to defend it when needed. A tool cannot do any of these things.

Data privacy is the point graduate students most often skip. Before uploading interview transcripts from a qualitative study, participants’ open-ended answers or data collected from students to an AI tool for analysis, you need to think very carefully. When your participants signed the consent form, they most likely did not know their data would be passed to a third-party service. If your ethics committee approval does not cover such processing, uploading this data to a tool may mean stepping outside what the committee approved. Even anonymisation is not always enough, because a detailed interview transcript can contain plenty of contextual detail that makes a participant recognisable.

Similar care is needed for your own unpublished work. And uploading a manuscript sent to you for review into an AI tool is a serious problem in terms of both confidentiality and reviewing ethics. Manuscripts in peer review are entrusted to you in confidence.

I also suggest asking yourself one question often: could I defend this text line by line in front of a committee, without the tool’s help? If the answer is yes, you are probably using the tool as an assistant. If the answer is no, the tool has started working in your place.

See you soon.

Image: Suhiresearch, CC BY-SA 4.0, Wikimedia Commons.