Should a student use GenAI to polish their text?

The instruction that it's ok to use GenAI to 'polish the text' seems pretty common in higher education research, from bachelors to doctoral level. I recently read this in an explanation of how someone used GenAI "I used AI tools, as is accepted academic practice, to check phrasing and suggest alternative wording." (I'm not providing the source for this because it was in the context of someone defending GenAI use and I don't want us to be distracted by this particular example).

I would dispute that it is accepted academic practice in all disciplines, but this depends on one's interpretation of checking and polishing. It is definitely accepted to use a thesaurus to suggest alternative word choices, for instance, and GenAI might offer this plus some help with choosing more common prepositions or verbs or context to make the alternative words sit better in the sentence. If you are writing in a second or third language, you might lack confidence in every single sentence. And as checking individual words is a tedious process, the temptation to do a block check and replace of the whole document is surely very, very, high. How many words need to be changed by the GenAI product for this to become more than checking and polishing? My experience in reviewing scientific papers in education is that it is very easy for a document to look polished but for paragraphs to lose their connection to one another, making the whole very difficult to read.

Here's what I suggest: the student and supervisor need to talk about what they mean by these kinds of vague phrases. Try out different approaches together with your real text, and discuss what happens to the text and if the final output is acceptable in your discipline area. Just having the conversation and practice will help you both to come up with a definition which works for your research, and will help the supervisor give more specific advice to the next student.

Reading the quote in Rachel's response, I am wondering what exactly was meant by "AI". Word has in-built spell and grammar checks, using which I would agree are accepted academic practice. Also tools like grammarly seem to be widely accepted. But while those use AI, they are not LLMs. For me, personally, acceptable "polishing" would mean to look at a text on a sentence-by-sentence basis and make individual decisions about whether a suggested word or phrase actually improves readability and the point I want to make -- or not, similarly to how one would work with the tools I mention above. Copying and pasting a whole suggested rewritten paragraph (or more!) from a LLM would, in my perception, not count as polishing any more but rather would be having a LLM rewrite. And there I worry that nuance (or even the meaning in general if I am trying to say something that isn't mainstream in the LLMs training data) is getting lost, and I also would hate to loose the feeling of ownership of my text. But I agree that this is a conversation that needs to be had between supervisor and supervisee in their specific context, and that probably also needs to be revisited as tools develop and with them our perception of what is acceptable and what isn't.

Let's think back a little, before 'automatic language 'correction' or 'improvement'' became an everyday possibility.

For years, I've not understood why people seem to not use spell checkers. True, this doesn't prevent grammar errors or homophone confusions (e.g., two, to, too; there, they're, their; break, brake), but it seems so simple, and changing language isn't difficult either (at least once one knows how).

I've never used grammar checkers in English, but as I write more in Swedish, I realise that I would benefit from having some sort of grammar checker there (because of the programs I use, it's not something that's easily available). But there's a problem: if a sentence, paragraph or document has mixed tenses, how will a computer know which tense is actually relevant for any given verb? (It's entirely possible that the relevant tense is not even present in the original sentence!)

In short: students should always use a spellcheckers. Grammar checkers should also be broadly encouraged (but, in students' first language(s), not considered 'compulsory' in the way of spellcheckers).

But these are 'conventional' tools. What about those that propose--or simply make--changes to the language that are in some way 'stylistic'?

I strongly dislike Word's language checker that, for example, I encounter in shared documents. For most professional purposes I write in my first language (English), and I am sensitive to nuances and the emphasis that comes from phrasings. I disagree with the majority of Word's language checker's 'suggestions'. To take three examples from a manuscript I am working on: "school teachers" is NOT one word, "appears to combine" should NOT be replaced with "combines" (we are making a judgement), "Perhaps the" should NOT be replaced with "The" (we are suggesting something). To follow these suggestions is to lose preciseness and, in the latter two cases, pervert the meaning of the sentences.

I do not know to what extent these suggestions are LLM driven. However, if I click 'ignore' in one editing session, I am met with the same suggestion the next time I go back--which is very irritating and distracting, and also adds considerable pressure to comply just to get rid of the nagging.

I sometimes think that students often feel that they 'must' use particular tools, even all the time. Therefore, from a teacher perspective, I would suggest making it explicit to students that while it's no bad thing to see what grammar (and language) suggestions their favourite word processing tool makes, this can be left until late in the writing process, and the bulk of the work done without distraction: turning off all grammar and language suggestions while writing is allowed! (This may actually help students who struggle with 'writers' block' since they are not being reminded to agonise about language while drafting, I don't know.)

Now, what about using generative AI for 'polishing', where a text is uploaded and 'improved'?

Here I start to doubt usefulness. As Rachel wrote, paragraphs and sections can start to lose their connectedness (I've seen this several times in clearly genAI 'peer'-reviews; it's one of the clearest markers). I have not tried putting in a coherent paragraph (or section) and then asking an LLM to 'polish', but it could be a useful exercise for both teachers and students. A very important point made by Mirjam is that, as in the conventional situation, suggestions should be marked--or changes explicitly tracked. Thus a rewritten sentence should look as it would in a word processer with tracked changes on. I do not know whether LLM programmes (can) do this.

There is also the problem of (lost) nuance described above. And the potential perversion of meaning and coherence, especially if done at the paragraph or higher level.

When it comes to argumentation, I have experience of peer reviews that are very nicely written 'thing commented on, reasoning, suggestion for authors', 'summarised' in as many words, to an almost dictatorial 'instruction to authors, justification'. It's inhuman and unpleasant to read, and I am pretty sure the 'summarising' was by generative AI. I do not know how universal generative AI's argument 'inversion' is, but it doesn't seem to be how to convince anyone. Logical and convincing arguments are made by leading from premises to conclusions through reasoning, not by post-hoc justification of a preposition.

Let me make one final comment on the fact that LLMs are 'next word predictors' and the critical importance of nuance. There is a HUGE difference between 'I really don't want to do something' and 'I don't really want to do something.' (The difference between the two statements is the order of the two words 'don't' and 'really'.) The former is a very strong preference: continuing to do the thing will be under duress (or close to), and the experience may be somewhere in the realm of unpleasant, highly distressing, panic-inducing, or traumatic. The latter statement is much milder, and continuing to do the thing will be OK; perhaps it will not be the most pleasant or enjoyable thing one could do, but nothing more than a transitory discomfort; a bit of relief when it's completed might be expected, but there would never have been any lasting harm. I do not know which is the most frequent ordering in English, but whichever it is, it is presumably the one an LLM will tend to give. And this means an LLM may either exaggerate mild discomfort, or completely invalidate potential trauma.

These are the warnings and dangers I see with using LLMs for 'polishing'. It's one thing to upload a single sentence 'I'm not happy with this sentence, can you suggest alternatives', but quite another to request 'polish within style x'.

One point I have not discussed, but is of course important, is what students learn by having to think about and make decisions about details themselves or don't learn by passing over the painful work of going through with a fine-toothed comb to a next-word predictor.

I would generally advocate to remind students that they can make choices about what sort of human they are, and that it is (arguably increasingly) important that they exert their personal human judgement wherever possible.

Writing may have several functions, two of which are (1) communication with others and (2) thinking -- which could be understood as communication with oneself.

To start with the easy part first, it seems fairly clear to me that your thesis advisor does not want you to short-circuit your use of writing as a way to develop your thinking. So don't do that.

The more nuanced part is the function of writing as communication with others. Here, there are multiple sub-functions. Again, let us highlight two of these: (1a) writing according to the 'genre' of your discipline, (1b) writing according to the rules of grammar and syntax of whatever language you are using.

Again, it seems fairly clear to me that your advisor does not want you to short-circuit practising writing according to the 'genre' of your discipline. In particular, it is hard to completely decouple this function of writing from the 'thinking' part, since adhering to, say, the genre standards of writing mathematical proof is very much aligned with adopting a 'mathematical' mindset.

So this leaves the use of generative AI to ensure that you write according to the rules of grammar and syntax. I assume this is what your advisor has in mind. However, again, it is not entirely obvious that this can always be decoupled from the other functions of writing. Indeed, if you have poor mastery of your discipline (or of anything else), it becomes hard to find elegant formulations. And when you do find those elegant formulations (yourself), this may be exactly what will trigger those "aha" moments that are the sign of new insights.

So, the bottom line is that, unfortunately, that there is no easy answer to your question. But I hope this attempt at addressing it helps.