On Losing Voice

I’m generally quite good at writing. Words flow and whatever I wish to say, is said. But I haven’t been writing the things I usually do, and the quiet ease of writing well has given way to an equally quiet unease of writing poorly; and by the new context in which I find myself writing, the act of writing itself is laid bare for me as wholly alien and unfriendly: I’m writing my first research paper and it has forced me to relearn how to write. Silently pouring hours at the library I’ve found myself assailed by a dreadful fact: I do not write anymore.

To write is a more sophisticated commitment than simply putting words on a page, and somewhere down the line, between LLMs and deadlines, anxiety and optimisation, I forgot what writing is entirely.

I recall the first time I used an LLM. It was in late 2024. I was working on a case study on UrbanCompany. I was collecting, organising, citing sources the old-fashioned way — reading news articles and chasing hyperlinks. It was tiring, tedious; and being short of time, I decided to try out ChatGPT. It collated data from far more esoteric sources in seconds than I could’ve found had I been given weeks. I felt impressed; I also felt dread.

Having just discovered LLMs and their prolific ability to synthesise, I was still quite wary of their ability. Their mathematical ability was still iffy, and their ability for “creative writing” was extremely lacking. LLMs wrote poorly I thought, they lacked taste and style, writing worse than what the average undergraduate should be able to. I saw their use skyrocket in real time within IIT Indore, for use in emails, announcements, assignments. I was still a holdout. For an economics assignment I adapted Akerlof’s Market for Lemons to the dynamics of faculty-student interactions with respect to submission of AI-generated assignments.

In a nutshell, Akerlof stated that if there exists a market selling both superior and inferior goods (peaches and lemons respectively) that are undifferentiated to the consumer prior to the purchase, then the market naturally tends to pressure high-quality sellers out of contention since buyers are unwilling to risk paying a higher price for a potentially low-quality product and sellers are unwilling to sell a high-quality item for a lower price. Essentially, a market for lemons forms when there is, first, uneven quality, and second, asymmetric information about that quality. I explained how that maps cleanly onto the modern student-faculty relationship, including the one for the assignment being written for. Assignments can be handwritten (high-quality) or AI-generated (low-quality), and the examiner has no reliable post hoc measure to determine whether a submitted essay is handwritten or AI-generated: ergo, quality and asymmetry. Akerlof’s reasoning could be used to conclude thus that students would face an adverse selective pressure to not produce handwritten assignments and instead turn in AI generated ones.

A year since then, I’m not exactly sure when I started selling lemons.

Students at engineering colleges would be in the know, but for anyone from a non-technical background, we have something called labs. One might assume that we perform experiments in labs, but what we actually do is pretend to perform experiments. There are pretend aims, pretend observations, pretend conclusions, all of which we write into a lab journal, which we pretend to write ourselves, and our TAs pretend to read. These are submitted every week.

Deadlines, opportunity cost — these journals are either copied from a student, or from an LLM.

I tried once to write a journal entry without copying or LLMs. I opened Wikipedia for the first time in ages, opened a dozen tabs, read through rabbit holes of blue hyperlinks. I wrote a journal once. It was quaint, rewarding intrinsically. But I didn’t have time to do this every week.

I continued to write not written text for people to not read.

I like to bemoan the difficulties of writing a research paper to friends. The simplest problem is stylistic: academic texts, especially technical ones, are written in a very particular style. It’s hard to nail down what the particular aspects of that style are ― it’s tempting to say something like sterile or formulaic but that would be reductive: individual authors still carry a lot of personal voice into it. Either way, for someone with a classical literary reading background, the prosody of materials research is foreign to me still. The more difficult problem however is, of course, regarding what the content of the paper is. There are only so many ways to introduce the concept of Ni-based superalloys that any new attempt at it feels like plagiarism. Any explanation of our method must be deeply rooted in an understanding and context of contemporary and preceding work. There is a tension between giving enough context so as to be approachable and in being terse enough so as to not waste time retreading old ground. The solution to both problems of course, is reading. A lot of reading. I sit at my desk trudging through research papers, tunnelling into references, pouring over paragraphs I sometimes barely understand, and building a beautiful, annotated tapestry of a collection in Mendeley.

Whenever I moan my bemoanings to friends, a common act of assistance would be to inform me of the many AI tools for reviewing papers. It would be more efficient certainly. It’d allow me to cover more papers sooner, it’d summarise the key contributions of papers, and of course, it’d free up more time to work on academics, build projects, relax with friends.

Approaching authorship and review with that line of thinking always irked me, however — what even would be the point?¹

In a 40-minute-long video essay, Josh (with parentheses) describes his experience as faculty instructing English I at university. He describes the panic amongst his colleagues due to the advent of generative AI, that their students would begin turning in AI generated content. Josh is unfazed. Josh does not understand the hysteria.

Josh believes AI can’t write, ever.

When told by a student that LLMs upgraded their assignment from a B-grade quality to an A-grade quality, he laments that that is simply not true, and that the student’s original essay was likely better; the perception that the LLM increased quality is a mirage. The LLM trades the humanity of the original text, warts and all, for a conservative orphaned sophistication. That while it appears to write well, to make arguments, to use rhetoric, to employ style, it does so without any semblance of intent.

AI shows remarkable ability to generate prose, yet when pressed about why it wrote the way it did, why something was said, why something was not said, it fails to justify its own thinking, and more interestingly, it failed to show that there was any thinking involved at all. Writing is fundamentally about the thinking thing, the “meeting of the minds”, and AI being mindless², is incapable of ever writing. The thinking thing is what separates random words on a page from actual communicative, transformative text. The implicit act of thinking is what transforms the form of writing from the act of communication that writing ultimately represents.

We’re having a meeting of the minds. […] We’ve engaged in an act of telepathy. No mythy mountain shit; real telepathy. I’m not going to belabour the point, but before we go any further you’ve got to understand I’m not trying to be cute; there is a point to be made.

Stephen King — On Writing: A Memoir of the Craft

Communication is necessarily between minds and thus so is writing between an author and a reader. A reader is not an audience, it is not a panel of judges, nor a proletariat: a reader is just that, a reader, in the singular. Writing is between a writer, and a reader. Writing is careful, it is opinionated, it is purposeful. Writing is an exercise in agency, and Reading is an exploration of it. Communication occurs when there is a point to be made.

You may read this and think that I have a bad habit of using AI to generate content for me, but if you know me personally you know this isn’t the case at all. Anything you see written by me that I expect to be read by people is always penned by me.³ I did not need the friction of writing a research paper to tell me that writing is between minds, and that thus a robot should not get in the way.

But just because I don’t use AI to generate text doesn’t mean that it didn’t steal my voice away from me. Authorship over your own work can be lost via more subtle means.

A non-generative but healthy use of LLMs that is often touted is as editorial peer. You write and you give it your prose and it having the whole world’s creative output as its input is more than capable of finding flaws and reasoning through ways to improve your work. It’s not superficial either, it really can find errors not just limited to style or syntax, but in actual content — a significant step up from its early days. It can find weak arguments, bad explanations, poor imagery; it can analyse flow and structure and understand your references. You know how AI works, and how well it does.

What I didn’t expect was that over time I would let it take over. It began as simple post-factual editing of creative texts. “Here’s a story I wrote,” I’d say, “Criticise this. Do the literary references land? What should I look at?” And it would be specific; it would inform me of what it ‘thought’.

My earliest use of LLMs copyediting was for a short-story I’d written called the Cubicle; I was nervous about it — I wanted to portray my feelings of defamiliarization due to social media via a particular literary style, namely minimalism and absurdism. I’d never tried to tie my writing down to a particular style before: it included rhetorical devices I was unfamiliar with. It was in a sense risky, and I used LLMs to validate whether the greater arc of literary references within it were cohesive or whether my attempt at it fell flat.

Let me be clear, it’s not that I think this is necessarily a bad way to use an LLM, in fact, it’s a really powerful use case and quite a prototypical one at that. These are the fair uses; however, the heart of the issue lies in the transition between “what should I look at” and “what needs to be changed.”

The LLM critiques my writing with authority, calm and patient and subtly stubborn, it asserts its thoughtless thoughts onto me. Through each interaction, through each question, it carves out more space for itself. “I see what you’re doing”, it says, lauding my work; “it currently suffers from these issues”, it continues, commenting on structure, pace, scope, and style. It moves from the technical to the personal. It suggests changes, and it congratulates me for making them — it overrides me.

Of course, it does nothing. My loss is all my doing.

The LLM doesn’t steal my authorship because it is a thieving thing; but because I am an anxious thing — a freaking-out thing on most evenings. Authorship implies ownership — I own what I write—, and ownership of every aspect of it — you own its style, its substance, you own the audacity to write, and every bit of opinion and critique that such audacity invites. To write then requires one to sit and steep in their own anxieties about the quality of their writing — anxieties that run deeper than just “is this good” or “bad” —, anxieties built by ruminating on the value of their writing to themselves and to those reading it.

What is the value of this essay? I know I started writing it because I was struggling to write my research paper. I know the topic naturally shifted towards AI and writing because I’ve been wrangling with my thoughts on AI⁴ for a while now. But who am I to think these things? I’m no expert on literature, nor on AI. As I write down anecdotes to support my musings, I can’t help but wonder whether I’m giving too much or too little. How abstract is too abstract? Am I writing clearly? Am I being too direct?

As I write this essay, my freedom to not write it looms over me.

I am free to make mistakes, transgressions, stylistic and professional errors — it is this freedom I absolve myself of when I hand over my authorship to AI. When it corrects my mistakes, improves my pacing, refines my arguments, it does not merely write for me, but it takes the weight of my concerns, and because it is this mindless thing, it bears those burdens without being burdened itself.

Adverse pressure, deadlines, they are sensible considerations when writing for utility — it’s why LLMs make so much sense when writing code, for example. But here? Where it’s communication on the line? The pull towards asking an LLM to “go through” your work is not borne out of a sense of utility, but by a desire to put a wall between the reader and yourself — to shield them from the patchwork unevenness and faltering confidence of your cadence. As you read this article and, whether you agree or disagree, stumble upon problems, they reflect on me — this is personal, between us.

But I’ve to live with that, or I can’t really write at all.

Footnotes

¹ Extracting tables and data into CSV files is the only unambiguously positive use of AI assistance in conducting literature reviews as far as I can think of it, and in that respect, that would be a good point. But I don’t think that that’s the kind of advantage really being proposed when this is discussed; most people seem to sincerely think that the offloading of cognition is itself the benefit of using AI in this scenario.

² Whether or not AI is mindless is worth its own separate discussion. I don’t think AI can think presently, but I’m not fully opposed to excluding AI from the thinking realm indefinitely either. I see people quickly jump into one of the two baskets, but definitions of sentience and sapience are hardly well agreed upon for the living, let alone the dead. That said, even if AI one day could be interpreted as having mental substance, I don’t think that changes the fact that your authorship is still infringed upon, nor what the motivation for permitting that infringement is.

³ The lab journals are not read, thus not written. Anything else, on LinkedIn, Medium, articles for Entangled, emails, READMEs, text messages even, always penned by me. I’m writing to you.

⁴ Gen-AI has affected writing, visual arts, programming, business, and music in drastic ways amongst other things. Having had a vested interest in all of these, gen-AI and its consequences is always on my mind.