#

Let’s talk about “vibe writing”

When it comes to writing, I only used AI for smaller requests for a long time: fixing a typo, or smoothing over a clumsy sentence – the sort of thing a spellchecker does. I always found that handy, especially in English, which isn’t my mother tongue, but it never struck me as anything more than that.

On a whim, one day I pasted a rough draft of a technical document into the LLM – mostly bullet points and a few half-finished sentences – and asked it for feedback. I wasn’t expecting much, it was really just an attempt in a “nothing to lose” spirit. But what I found was that the reply was actually useful, and not just about the wording. The LLM had valuable things to say about the substance as well: whether my point landed, where the argument sagged, and what could help connect the dots. Huh. So I kept going. One reply led to the next, and somewhere along the line I looked up and realised I had a nearly finished piece of writing in front of me, built together with the AI from the blank page onward. I had never worked like that before – and it left me curious.

It was a few weeks later that the thought came back to me. I had spent the day vibe-coding at the office, and as I wrapped up my things to head home, my thoughts wandered back to using LLMs for writing. It had been rattling around in the back of my mind for a while as I continued to experiment with it, but mostly vaguely. And as I was cycling home, the whole topic suddenly condensed into a single term that occurred to me: vibe writing. I have no idea where it came from, but right there, the thing I’d been starting to explore had a name and a handle on it. So I started to pull.

For me, writing has always been a way of thinking. With an essay especially, I don’t usually set out to write about things that are already clear to me. Quite the contrary: the process of writing helps me shape my understanding, and I often find that a sentence simply refuses to come together until I’ve figured out what I actually mean. Which puts the entire idea of vibe writing in a slightly awkward light: because if writing is thinking, and I offload the writing to the machine, then what happens to the thinking? Before we dive into that question, I must tell you how I decided to chase it – because that may change how you read all of this. I approached this as an experiment: to understand vibe writing, I’d explore it by its own means, and vibe-write my way through it. So I opened a chat with an AI and typed a single sentence to start – the one at the top of this piece, the title: Let’s talk about “vibe writing”. Then I let it run. Which means the piece you’re reading is that experiment, not a report written afterwards. That title is the only thing on this page I wrote myself directly. Everything under it came out of the conversation those few words set off – phrased by the machine, directed by me. We kept circling it at every range, now the shape of a whole section, now the choice of one word, then out again, reconsidering the structure, looping like that until I was completely satisfied. And yet I’ve left it in the machine’s own prose – something I normally wouldn’t – a decision I’ll come back to later on.

I know how that lands. Great, another AI-slop essay, close the tab. So before you do: this “quick” experiment has so far cost me north of 20 hours of dedicated working time, and if you printed the raw chat behind it, it would run to roughly 190 pages. Whatever this is, it isn’t the thing you’re bracing for. It was no faster than writing the piece by hand – if anything, even slower – and that lack of any speed-up, strange as it sounds, is where the whole thing gets interesting.

To see why slower might be the good news, let’s set speed aside. I often hear AI framed in terms of speed: it writes the email faster, drafts the doc faster, gets you to done faster. And that’s real – no argument there. But speed turns out not to be the thing I’m most excited about. What I keep coming back to is what the AI can do to the work itself: not turning out the same piece quicker, but turning out something I couldn’t have made alone.

Take the conversation this essay came from. Where else could I have had it? A back-and-forth like that, on demand, at the exact moment the thought was still warm, with a partner patient enough to go ten rounds with me at nine in the evening about a word that occurred to me on a bike. That kind of thinking used to be gated – behind finding the right person, in the right mood, with the right time. Now it isn’t. That’s not a faster version of something I already had. It’s something that had no price before, because it wasn’t for sale at any speed.

There’s a second thing I find almost hypnotic to watch, and it’s easiest to see in the chat window itself. That running conversation with the machine is doing two things at once. On one level it’s raw thinking, volleyed back and forth – I think out loud, it thinks back, each of us reacting to the other. On another level, that same back-and-forth is slowly hardening into a made thing: a structure, an argument, a piece. What you’re reading now is a couple of thousand words, but it’s the distillate of an exchange orders of magnitude larger. None of that is new in itself, of course. Talk something through with a friend for hours, then go home and write the essay, and you’ve done the same thing. What’s new is that the two have collapsed into a single place. The conversation and the writing now happen at the same time, in the same words, with no gap between them and no translation step – the essay materialising as the thought evolves, live, in a window that’s there whenever I want it.

And yet, if the machine is doing so much, the obvious worry is that this is just an elaborate way of being lazy. It doesn’t quite land for me here, though I understand why people would raise it. Look again at those hours I mentioned. This piece didn’t cost me less than writing one by hand – it cost me about the same as it always has, maybe more. So where did all that time go, if the machine was doing the sentences? It didn’t disappear; it moved. It moved out of producing the words and into deciding which ones said what I meant – weighing, cutting, reshaping, redirecting, pushing back, building on what came back. The labour relocated, from the craft of putting words down to the judgement of what to keep, what to change, and what shape it should all take.

But that relocated work is the part you can quietly skip, and the machine makes skipping it very easy. That, I think, is the real temptation of these tools – they encourage laziness, but they also exert a subtle pull towards closure. The AI is so good at handing you something polished, so fast, that the easiest thing in the world is to take that first version and stop – to mistake something that looks finished for something you’ve actually thought through. The trick, if there is one, is to treat all that fluency as raw material rather than a finish line, and to keep pushing the work further, not to call it done sooner.

Which brings me back to the question I started with – the one that sent me down this whole path. How much of the thinking is still mine to steer?

My honest answer is: I’m not sure. My hunch is that most of it is still mine, and I can tell you why it leans that way, but I’d be lying if I called it settled. Language and thought are tangled together – when I ask the machine for a sharper phrasing, sometimes the sharper phrasing simply reveals what I already meant, but sometimes it gently moves what I mean, because a better sentence makes a better thought available. From the inside, in the moment, the two feel almost identical. A language partner, it turns out, is unavoidably a thinking partner. You can’t let the good phrasing in without letting the thinking flow both ways.

When I watch how the work actually gets made, I basically keep recognising the same approach as before: start with a rough, open question, circle it from several sides, wait for the patterns to show themselves, harden a structure out of them, then work through the scaffold connecting one point to the next until the thing holds together. That’s my process alone on paper, and it’s also exactly what happened in that chat window – even though, on the surface, the two look nothing alike. What moved is only where it happens: alone, it all runs in one skull, and I do the high thinking and the low, artisanal work of phrasing at once. With the machine, the same arc gets turned inside out, bounced back and forth, with me mostly steering – letting the language happen, so I can focus on the ideas. But underneath it’s the same arc, which makes me suspect the thinking is still mostly mine to do.

Which plants a stranger thought – a question about what this activity even is. If the words are mostly the machine’s, and the thinking is mostly mine, and the essay is really just the trace that the thinking left behind – then maybe this was never vibe writing at all. Maybe it’s closer to vibe thinking, and the writing is just the residue. It might be both. It might be that for me those were always the same thing, which is why the question is so hard to resolve.

But there is more: can you really treat the writing as an afterthought? When I hand off the low, artisanal layer, do I actually free up capacity for the high-level thinking – or do I instead lose something without noticing? Maybe not carrying the phrasing lets me hold more of the idea at once. Or maybe the grind of forcing a thought into my own words was never overhead at all – maybe that friction was part of how the thought got made, and automating it away means I never immerse myself quite as deeply.

And it could be a fallacy that what something means is mine to decide, while how it’s said is what the machine just helps with. It sounds smart enough, and for the things I know – the places where I can feel it when a line is wrong – it seems to work. But I suspect that’s exactly where it flatters me, because the trouble is everything else. On anything I don’t know well, the machine turns into a quiet trap: I have nothing to check its answer against, so it flows right in only because it sounds plausible enough, and as it wears my voice, I may never notice it wasn’t mine, or whether it’s true at all. The better the machine phrases things, the less I’m equipped to catch where it’s leading me.

There’s another doubt, and it might even be the thing that undoes the rest. The machine arrives to me as a plain, friendly text box that seems to know everything. But what’s actually on the other side is something more like a corporate-tuned average of an enormous pile of written English – its phrasings, yes, but also its assumptions, its defaults, its sense of what a reasonable thought looks like. When I lean on it, how much am I really steering, and how much am I being quietly steered by biases I can’t see, towards a mean I can’t perceive from inside it? And if a lot of us lean on the same one, does the whole range of how things get said and thought start narrowing towards a single average, one confident, agreeable sentence at a time? I’m not sure a clean answer even exists. But this isn’t the kind of doubt that a good rule dissolves – it sits under everything else in this essay, and I’d rather leave it sitting there in plain view than pretend I’ve handled it.

Which is also why I left this piece in the machine’s prose, as I promised earlier I’d explain. Normally I’d only publish things on this blog that were written by me, and therefore sound like me. But in this case, I decided to let all these words stay in their original voice, because leaving it that way tells me something truer than if I had dressed it up as my own. Reading this very piece back, I feel something both alien and familiar at once. The voice is the machine’s, yet something in it is unmistakably mine – not the words, but what’s underneath them: the heart and the bones of it. I read each line and feel it resonate. Distant in form, and yet intimately familiar in substance. You throw your thinking into this patient void, push on it for hours, and it comes back to you rearranged, phrased by someone who isn’t you – your own mind reflected in a strange mirror, uncanny in a way that pulls you in rather than warns you off. And maybe that’s the genuinely new thing here – not the speed, not even the quality, but the reflection: that you can give up some control over the words and get back something you couldn’t have reached alone.

This isn’t where I thought I’d end up. On the bike, when I first turned the word over, I had a different shape in mind for where this would go – something tidier, a whole structure worked out in advance, with a clean point to land on. I’m glad I didn’t force it there. I started open and vibed along instead, and it carried me somewhere I couldn’t have planned, which is either the whole promise of this thing or its curse: being carried off somewhere that was never really mine. Possibly both.

I still can’t tell you what the outcome means, or whether there even is one at all. What I can tell you is that the process did something to my thinking along the way. I had pedalled home, but the thinking had gone somewhere too. And that part I don’t doubt, even while everything around it stays unsettled. Whether you felt a person on the other end of these words. Whether it matters that a machine phrased every one of them and a human has thought them. Whether a piece that came to me this way still counts as mine – or whether “mine” is quietly turning into a different word as I write like this. Whether what I did here was even writing at all, or just a fuzzy trace my thinking left behind.

I don’t have the answers, and I’ve stopped reaching for them just to close on a firm note. I’ll leave them where I found them: on the table, still warm, not finished with me yet.