If you haven’t seen the news lately, then you may not realise that AI has taken over the world.1 Everyone’s using it, everyone’s talking about it, and everyone is being replaced by it. It is simultaneously increasing the number of jobs and eliminating them or having a tiny effect on jobs.2 It is simultaneously going to be the golden ticket to early retirements/short work weeks with other pleasures of a secure life, and the cataclysm that will destroy us all.
So, if you haven’t heard about the wonders of AI, you are worse than the troll that lives under the bridge. You are the bridge that the troll lives under!
How can you not worry about it? How can you not rejoice in it? How can you not fear it? How can you not love it? How can you not talk about it?
I know one thing… the what-now-looks-like-an-inevitable-crash is going to doom retirement funds, while simultaneously making the rich richer (comparatively speaking, of course).3 So, business as usual.
What I wanted to talk about was something much more prosaic: What do departments do in such a fast paced and breathless environment? Apart from asphyxiate, of course.
While what I say applies to undergraduate education too, I mainly have graduate education in mind here.
I am going to suggest a very boring answer. Just as we don’t let developing students use calculators, we don’t let (graduate) students use LLMs for writing or idea generation or analysis for coursework and qualifying papers, because our job is to develop them along these very lines and LLMs prevent them from doing so. Furthermore, more research is not the goal; well-trained and thoughtful researchers are the goal.
Let’s note some important reminders, so that we feel like we really care, and then proceed to ignore them. First, LLMs are still terrible for our environment and remain a serious threat to the United Nations Sustainable Development Goals. I recognise that worries about sustainable development are rather quaint in the era of LLMs and (hot again) gas vehicles, but global warming is still the true existential threat in front of us. Maybe, total LLM usage will have a sustainable footprint in the future, but we won’t get to that future if we keep doing what we are doing now. But, I guess most universities, researchers, and students who use LLMs have accepted “who needs water or a world to live it later?”
Second, our electricity bills are going up thanks to all of our LLM usage and the consequent establishment of data centers; folks who live close to such data centres have much worse health outcomes already. Yes, they live far away from me, so it’s doesn’t feel so real, but my usage contributes to it. I recognise there are many such trade offs in life, and we often can’t control many such decisions, but, in this particular case, because LLMs aren’t unavoidable elements of our workflow yet, it is important to think about whether such usage is more important than someone’s health.
Third, I am not here to say that LLMs are not useful. Weedkillers are very useful too, but it is clear we shouldn’t be using them when we consider everything else they result in. Fertilisers are an even more interesting case — there is no doubt (unlike with AI) that they have completely revolutionised our lives, and even here we have realised how terrible our decision of profligate use has been. With LLMs, it is the economic warnings of Nobel Laureate, Daron Acemoglu, that are slowly but surely coming true, and not really the promises of a golden future that the tech community has promised.
We can hide our heads in the sand, but each of us using LLMs contributes to these deleterious effects in real ways.
OK, the self-flaggellation is complete. Like good folks, we have acknowledged our sins; and now, like even better folks, we can act like it doesn’t really matter in our decision-making and move on.
I have used LLMs (the free ChatGPT version) and killed a few trees and hurt humanity’s future and others’ health for mundane admin work (“Bullshit jobs” in the evocative words of David Graeber) and as a search tool a few times. I also used LLMs once to create an app to convert guitar tabs into those for a baritone ukulele (it was surprisingly good for something that I spent all of 20 minutes on).
However, I don’t ever plan to use it for my research — either for writing or thinking. My reasons are two-fold. First, I like doing all those things on my own. It’s that self-centred way of doing my research that I enjoy the most.4 I don’t do research for others; I don’t write papers for others. I do it purely for myself, so I can be proud of the stuff I do, and because I genuinely believe no one else thinks my work is useful or good anyway. This was also true before tenure. But, I won’t lie — before tenure, I did have more of an eye on doing things that the field cared about at least a bit while indulging myself.
Second, I am too scared that if I were to use LLMs for writing and thinking, I would lose my already minimal skill-set. As with any skill, my research and writing skills stay honed only if I constantly practice and work on them. Giving the task to someone or something else feels like believing that one can remain good at playing tennis or the guitar by simply watching someone else practice or by holding a part of the guitar or tennis racquet while someone else holds another part of it.
Note, I am not talking about doing better research with/without LLMs. Even if one could do better research, it doesn’t help with developing my own skillset and therefore it doesn’t aid in my own personal development. My view is inward here and not outward, if that makes sense. Your mileage may vary on such considerations.
There is some interesting recent research on how LLMs affect users. First, with respect to writing, there is recent work showing that writing becomes more stale and more uniform as students use LLMs (no surprise there). But, the issue about writing that doesn’t seem to get as much attention is that it shapes thinking too. The multiple rewrites/edits help the author (at least, me) understand what I want to say, what my main point is, and what my best argument is. In fact, I experienced the very same questions while writing this very post! I wrote the following text, which I had always thought to be a good argument, and then realised when I wrote it down that there was no clear argument there:
The challenge is not in coming up with something new. The challenge is to do it in a way that sustains itself over extended periods of time. LLMs are currently able to come up with ‘new’ stuff as they have been trained on centuries of human thinking. But, what happens when the intuitions that reside in those texts are used up? How do we break through from past thinking or past frameworks? There is some evidence in the math results that LLMs are capable of such truly creative thought (to some extent), but I have not seen it in Linguistics (minimally, yet).
I ended up removing it as part of my argument only after writing it down and thinking more about it. Others might see more/less purchase with the argument. However, my point is that writing is not passive — it is an extended process of analysing and evaluating one’s own thoughts. Something that is lost if it’s pawned off to an LLM.
Second, we need to give our students opportunities to learn and grow, and that involves all kinds of trials and errors in their own heads (without any help from anyone). Getting used to wrong-turns/failures and being able to critique their own reasoning are extremely valuable skills. How do they learn such skills if we as teachers don’t provide the appropriate environment for it? They also need to learn to persevere when there is no overt evidence of success. These are the kinds of things that are likely to be lost if we outsource writing and thinking to LLMs. Some recent research has argued that the desire to persevere is weakened with LLM usage (at least in some controlled, short-term, experimental contexts.)
Until recently, I myself have maintained the opinion that there is simply no point to talking to students about this, even if I personally don’t want/like to use it for research. Students are using LLMs already, and no amount of advice and discussion about the dangers is going to change their behaviour. Much like folks are distinctly aware of how bad social media is for their mental health, but can’t stop themselves from using it, because the fear of missing out is a real problem.
However, I think that way of thinking about the issue is a bit defeatist. Yes, folks use calculators, and all kinds of technology in their work place; however, we still don’t let students learning about arithmetic use such equipment, and rightly so. Our responsibility as teachers is to ensure that our students are capable thinkers and researchers, and not let them develop crutches. And sorry, but getting ideas from LLMs and using LLMs to write papers doesn’t cut it.
So, the more I think about it, the more convinced I am that, not only for environmental reasons, but also reasons of pedagogy, we must insist that our students not use LLMs for writing or idea generation or analysis for coursework and qualifying papers (and, really, most research tasks related to those things, except perhaps searching for relevant papers). Minimally, this is a conversation I will be having with my own students. I will also be making a recommendation to my graduate linguistics programme in upcoming conversations about LLM use that we should absolutely be clear and prohibit LLM usage for writing or idea generation or analysis for coursework and qualifying papers — not because we can monitor such usage, but because it sets clear expectations.
Note, I have said nothing about what I’d recommend with the dissertation or with journal submissions. In the case of dissertation work, I just need to think more about it — although, currently, I lean towards saying no use of LLMs in that case too (qualified in the same way as above), the reason for my reluctance to commit to that recommendation is that perhaps students are sufficiently developed by that point to use LLMs in a mature way. In the case of journal submissions, that’s their business and involves a conversation with their collaborators, and I have nothing to say about it.
I now believe that the attitude of either let’s-discuss-it-but-otherwise-wait-and-watch or let-students-do-what-they-will-anyway doesn’t achieve anything productive, but it is an abdication of our resposibility as educators. Given how fast things are evolving, I might very well change my opinion in the coming years. I would be stupid not to minimally re-evaluate in a few years. However, it is important to draw clear lines when things are complicated. The line might be in the wrong place, but you only realise it’s so if you draw it in the first place. As far as I can see, not drawing a clear line is worse.
This is an actual prompt I just gave ChatGPT in a fresh unsigned-in chat session after I mentioned that I am an academic: “is the world better without LLMs?”
It gave me a long answer with plusses and minuses, and then I gave the prompt: “Well, choose the best side”.
And here is the answer I got:
The world would be better without LLMs. If I have to choose one side, that’s my answer.
Not because LLMs are useless, but because I suspect their aggregate costs to human intellectual life may outweigh their benefits.
They make generating information easier without necessarily making us wiser. They lower the cost of producing plausible arguments, but not the cost of determining whether those arguments are true. They can save researchers time while encouraging institutions to demand more output. And they risk turning activities that develop human judgment—struggling with difficult problems, writing through uncertainty, and learning to recognize one’s own ignorance—into tasks to be outsourced.
My biggest concern is that we may optimize the production of knowledge-like artifacts at the expense of the actual pursuit of knowledge.
That is especially troubling in academia, where the distinction between appearing knowledgeable and understanding something is fundamental.
I acknowledge the counterargument: LLMs can democratize expertise, improve accessibility, accelerate some discoveries, and help people who otherwise lack support. Those benefits are real.
But if forced to choose between a world with these tools and a world without them, I would choose the latter—not because I know that the harms outweigh the benefits, but because the harms could undermine the very human capacities we need to evaluate whether the technology is helping us.
My position is a judgment under uncertainty, not an established empirical conclusion. We haven’t yet seen the full long-term consequences of widespread adoption.
And I would apply the same skepticism to my own existence: an LLM shouldn’t get a presumption of being good for humanity simply because it can produce a persuasive argument for its usefulness.
Ironically, I couldn’t have said it better.
1. Students who already use LLMs will continue to use it even if you say no.
Hey, this was my argument, previously! However, I don’t think that is correct. Most students want to follow appropriate procedures. Students can currently pay someone else somewhere to come up with the ideas and even write. And yes, there may be some who actually do that. But, more importantly, there are many, many more who don’t use such services, and equally importantly, I don’t think I have heard of even a single graduate case in my university or elsewhere that has done that. So, if we are clear about LLMs not being acceptable, I don’t believe students will violate that expectation.
2. Aren’t L2 speakers disadvantaged by this, when it comes to writing?
I will speak to my own experience. The idea that one is a good writer just because they are an L1 speaker doesn’t accord with my experience. As I mentioned, to me, research writing is about organising thoughts and marshalling arguments, and poor writing, in my experience, is much more commonly associated with a lack of clarity on one’s own claims and arguments. Control over the language is the least of the worries, and one that is easiest to solve.
3. LLMs are inevitable. Don’t be a Luddite!
No, LLMs are not inevitable. Yes, silicon valley benefits a lot if we all believe that, particularly given they have sunk trillions of dollars on a technology that has produced little real improvement in productivity so far. Let’s not fall for the hype. There is a choice, and as long as we are humans, we get to make it.
Also, the Luddites were protesting worsening labour conditions. They were not opposed to the use of machinery in general.
4. Isn’t there a more moderate strategy? Maybe we can teach students how LLMs can make errors and then they can use that knowledge while using LLMs to do research or write?
Of course, we should absolutely be teaching students how to identify LLM slop and poor argumentation. There’s no two things about it. So, no disagreement there. However, the question of interest in this blog post is one of whether they should use LLMs in their own work.
If they use LLMs in their own work, we are back to the original problem. We want students to develop a deep understanding of their domain. Both that understanding and the use of that understanding are skills like any other. We don’t learn to paint well no matter how many paint-by-numbers we practice, and no matter how much feedback we get while we paint-by-numbers.
Similarly, we don’t learn to play tennis/guitar at a professional level no matter how well we play videogame tennis or Guitar Hero. We need relevant focussed practice — that’s the bottomline. The more time we have practising our skills, the deeper our understanding and the better our writing.
In fact, even after we develop an expertise, if we depend on crutches unthinkingly, our skills will weaken. There’s no avoiding that, as far as I can see.
5. How are LLMs different from calculators or spell-checkers?
LLMs are not like calculators. They don’t just perform a calculation that you have specifically asked them to do and that you understand how to do.
LLMs are not like spell-checkers. We may occasionally (often, in the case of some folks) not know the relevant spelling, but we completely understand the concept of spelling, and how to find out the spelling from other sources if we wanted to. Again, the crucial underlying knowledge is already present.
So, comparing LLM usage to calculators or spell-checkers is not appropriate. LLMs are a mix between social media and the person in Searle’s Chinese room. They have the power to convince you that they are real and they even sound human — much like how so many folks are tricked when they see all the happy photos of their friends and acquantances on social media and think ‘’Oh man! Everyone except me is having a fun life’’, LLMs have the power to convince you that they are real and even sound human; consequently, our sensibilities and our usual guardrails are easily tricked.
And like the person in Searle’s Chinese room, LLMs have no understanding or no real experience with what they are outputting. So, if students depend and collaborate with them while developing their research skills, there is a good chance that they too become no different from the person in Searle’s Chinese room — one who can go through the motions perfectly, but have no deep understanding or knowledge.
This post was conceived, written, proofread and even only read by AI. #NoHumanInTheLoop. [Edit: I was trying to be funny, but I realised it can confuse a reader. Good thing, comedy isn’t my day job.]
OK, we should really keep LLMs and AI separate, but I will use them interchangeable for the purposes of this post.↩︎
I highly recommend Dean Baker as a source of economics news, instead of the quite useless and hype-filled finance media. He appears to be a quite level-headed economist, and is a clear/honest communicator.↩︎
Also look at this article on the IMF’s recent report on the AI bubble.↩︎
I also don’t let others eat the ice-cream I bought for myself; OK, except, maybe my daughters.↩︎
For attribution, please cite this work as
Durvasula (2026, Oct. 19). Karthik Durvasula: [ɑɪ] don't use it for research. Retrieved from https://karthikdurvasula.gitlab.io/posts/2026-10-19- [ɑɪ] don't use it for work/
BibTeX citation
@misc{durvasula2026[ɑɪ],
author = {Durvasula, Karthik},
title = {Karthik Durvasula: [ɑɪ] don't use it for research},
url = {https://karthikdurvasula.gitlab.io/posts/2026-10-19- [ɑɪ] don't use it for work/},
year = {2026}
}