I Know What My Sentences Want
On AI, authorship, and getting told to write it yourself.
I use AI when I write. That's the declaration. It's not a confession, and it's not an apology, and I'm not going to dress it as either. It's a description of a practice I find genuinely interesting, and this piece is my attempt to say why.
(Technically... I don't actually use it all the time, in fact, far less than you might think – all of the en dashes you'll see in my writing are mine alone, for instance, but I wrote that sentence above 6 times and every time I tried to couch it, it didn't land as hard, and I'm a sucker for a good entrance.)
Before we go any further – I know the internet is full of machine-generated slop, and I know you've been asked to trust writers who've given you less and less reason to. I can't fix that. What I can do is be honest about what I'm doing and let the work make the case over time. That's the only move available, and I'm taking it.
I've been writing since I was old enough to know that arranging words was something you could do on purpose. Long enough that I have opinions about it, the way somebody has opinions about a city they grew up in: specific and impossible to be neutral about.
I'm left-handed, which means I came to writing through a body that had to solve problems before it could create prose. The smudging, the claw grip, the page turned oddly to stop hand cramps before a thought was transferred fully. Writing, for me, has always been a physical relationship with resistance, a set of small negotiations between what I'm trying to say and what the act of saying it entails. I didn't choose that, but I'm not sure I'd trade it. There's something in having to work around a thing that keeps you honest about what the other thing requires.
When the physical act of writing demands that much negotiation, you start noticing things: what a sentence is actually doing versus what it thinks it's doing, where the thought is real and where it's performing realness. You develop opinions about your own work the way you develop opinions about anything you've had to fight with long enough. I'm not saying being left-handed makes you a better writer, just more observant, because of the resistance curve...
Also, you have scissors and cutting things neatly – leave us with something.
All that to say – I know what my sentences want to do. I know where I reach for a comma because I'm being cowardly and where I reach for one because the pause is real. I know the words I overuse and the constructions I fall back on when I'm tired, and the register that feels like performance and the register that doesn't.
I know what my voice sounds like. That matters here, and I'll come back to it.
(I want to pause here to say I did, in fact, go on a triade at this juncture about how Grammarly is sucking the soul out of my writing and stealing my misplaced commas and smoothing my anarchic register switches, much more so than AI ever attempts to, but that digression, as entertaining as it was to write, was missing the point, so it is gone. Perhaps a different musing for a different day. For now, I will follow what the red lines say... but I am not happy about it.)
Writing is a particular kind of dualistic cognitive work where you're generating and evaluating simultaneously, and the evaluation always slightly poisons the generation (in the non-AI sense) because your critical ear is running in the same room where the music is being made.
What I find AI good for is separating those two things. I generate through it, then I evaluate against everything I know about my own voice, and the gap between what it produces and what I'll accept is where the work actually happens. Six hours on the specific texture of a single point. An emphasised correction so oblique that no AI could have written it. A paragraph taken apart and rebuilt four times, not because the first version was wrong, but because it was sitting at the wrong angle, and close is not the same as right. Close is where you stop if you don't know the difference, and the difference, I find, is the heart of it all.
It's not faster, frankly. It's slower, at least for me, when compared to not using it. The gap is not a problem to be solved. It is where the sum of both parts becomes a newer and more interesting whole.
I'm building a project called Analecta Hiberniae – found manuscripts in the Irish Mythological cycle style, dealing with real modern problems through that lens. They don't need to be written in paratactic structure. They don't need the monk's Latin, approximate and period-specific rather than textbook correct. They could be fully modern renditions.
They would be easier that way, and they would be less interesting by exactly the same amount.
Writing in a form that resists you, that has its own logic you have to learn and honour before it'll do what you want, produces things that couldn't have arrived any other way. The difficulty isn't incidental. It's generative. Working with AI is the same move for the same reason.
The voice is a set of resistances as much as a set of tendencies.
What I resist is mine. The AI doesn't know what I resist until I tell it, and even then, it only knows approximately, and I have to keep telling it, keep correcting, keep pulling the thing toward the shape of the thought as it actually sits in me rather than the shape of a thought that resembles it. That process is not nothing. I would go further: that is the job. It is the same job it has always been. The mechanism is different. The standard isn't.
What you read here has been through that process. Shaped, argued with, sent back, taken apart, rebuilt and added to – wholesale in many places. Not because I couldn't read what came first, but because what came first wasn't quite right, and I know what quite right feels like because I've been building that knowledge for long enough that it's the thing I trust most about my own practice.
I type some of the words, and it types some of the words. I decide all of them.
The honest version of authorship has always been messier than the myth of it. Writers have had researchers and editors and collaborators and arguments with themselves that lasted years before a sentence was settled. What's changed is the surface the argument happens on. What hasn't changed is that someone has to know what the sentence is supposed to do and refuse to stop until it does it.
That's me. That's always been me. That's not changing.
I had no better place to put this edit because just as I'm a sucker for a good opening, I like when a musing cleanly resolves to a point, and the constraint argument is a true and honest point that really does interest me – it forms the basis of why I use these tools at all. They are slower and, at times, more clumsy, but they are fun and interesting to fight with.
The reason for this insert, however, is not to double down on that point; it is to quickly mention that there are places where some of my writing will have zero AI involvement. Not because of any principled position, but because sometimes, the thing I am writing is so on the edge of madness that it breaks the AI and it asks me to go away.
I had a whole multi-part section on this, but it was moving away from the neat ending and is honestly far funnier to explore in detail as a first-class post itself.
That is not here today, but let me leave you with this fragment from Claude's exported brief I asked for after it got frustrated with me and suggested I just write the essay, a different essay, myself... Note that the "writer" Claude is referring to is itself.
What this session tried and why it failed
Multiple attempts were made to rewrite this piece from the ground up. Each one inherited the structure of the previous version — mythology, wanting, building, seven months, frames, Badb and ogham, five hundred listens, ending. The chronological order became the only imaginable order. The accumulated weight of the sessions made genuine structural freedom impossible.
The brief says: the five hundred listens might be the opening, not the ending. Everything else arrives as the answer to how that song came to exist. This was identified as a possibly truer structure, but never attempted because the writer had been, as he put it, poisoned with the truth, the same way Nora was. He knows too much to find the clean room.
This brief exists because of that. A fresh instance without the accumulated sessions is the only way to find out if there's a genuinely different structure available. Or whether the inherited structure is, in fact, the honest one — a piece about layers of construction that is itself a layer of construction and knows it.
That question is yours to answer. Not mine.
Beautiful.
The Analecta Hiberniae series is available now. If the constraint argument interests you, it will be a better demonstration of it than anything written here.