I write a fair amount and I care about the quality of writing. At work, where we write papers, submissions, I put a lot of thought into writing simply, concisely, and persuasively. Outside work, I wrote tons of social science papers, and nowadays I mostly write blog posts of me thinking out loud.
Over the past year or so, something shifted in how I write. The change went unnoticed. This is me trying to work through what changed.
Everything is pivotal
I find AI writing exhausting to read because it treats everything as life and death, Armageddon style. A tutorial about tweaking Claude Code, about B2B sales insights on LinkedIn gets the same weight as an essay about the meaning of life. Every paragraph reaches for significance, every transition connects to something larger, and the cumulative effect is like being lectured by someone who genuinely believes every sentence they’ve ever produced is the most important one.
The vocabulary gives it away first. “Pivotal”, “groundbreaking”, “transformative”, “robust”, “seamless”, “foster”, “delve”. These words appear so often in AI output that they’ve become huge signs to the reader that an LLM wrote this, or at least that the writer didn’t care enough to remove them. There are people maintaining actual banned word lists now, which is funny but also a bit sad.
But the bigger issue is the tone itself. AI prose inflates everything. It can’t just describe what happened, it has to frame what happened as part of a bigger story, as a moment of transformation. That register works maybe ten percent of the time. The rest of the time it reads like a press release for something nobody asked about.
The writing I used to do
I used to write things like this:
It could have been a one-off. But it wasn’t, unfortunately.
Two short sentences. The first one sets up an out, then the second closes it. I liked that move. I used it often, and I thought it worked well for the kind of writing I do, which is mostly personal, mostly honest, mostly just documenting what happened without pretending it was more dramatic than it was.
I also used en dashes, usually when I draft in Word or Outlook (works if there’s a dash + space + a character + space). Not constantly, but when a thought needed a parenthetical that felt too important for brackets and too casual for a new sentence. I varied my sentence lengths deliberately, mixing longer ones that developed an idea with shorter ones that landed because of the contrast. None of these are unusual techniques, and I’m not claiming they made me some great stylist, but they were mine, and I’d developed them over years of writing in a way that felt natural to me.
I’ve had to stop doing most of this.
The short fragment pairs read as AI now. En and em dashes read as AI. The deliberate sentence length variation reads as AI, because AI does all of these things constantly, without earning them. A short sentence lands when the long ones before it build the pressure. AI just alternates lengths mechanically, and the effect is that readers now pattern-match any variation as synthetic, regardless of who actually wrote it.
So I’ve shifted to commas. More connectors like “and”, “so”, “then”, “because”, “which meant”. Longer sentences that stitch thoughts together rather than letting them stand as fragments. My writing is more continuous now and less punchy. I think it reads worse for it, but at least it doesn’t immediately trigger the “AI wrote this” assumption.
You don’t notice until you do
I shifted my style guide gradually through small adjustments, swapping a dash for a comma here, adding a connecting phrase there, smoothing out a fragment into a clause. Each individual change was minor and they all felt like natural choices at the time.
Then at some point I noticed what I’d been doing, and once I noticed I couldn’t unsee it. It’s the exact same phenomenon as the GitHub Copilot pause in IDEs. Every developer who uses it has that half-second hesitation where they stop typing and wait for the suggestion to appear.
Most people don’t register it until someone points it out. But once you see it, you catch it in every pair session, every screen share, every time someone’s cursor lingers a beat too long after a line break.
This is the same thing but for prose. I now catch myself checking whether a sentence sounds like something Claude would produce, and that filter runs before almost every sentence I write, automatically, whether I want it to or not. That filter didn’t exist two years ago.
Writing defensively
My first instinct when writing a sentence now is to check whether it sounds like AI output. Only after that do I think about whether the sentence actually works. That’s a genuinely strange way to write, and it means the tool I use to help me write has also become the aesthetic I’m writing against.
The specific moves look like this: more commas where dashes or periods used to go, more subordinate clauses to keep sentences flowing, more informal connectors to stitch paragraphs together, and generally more visible effort to sound like a person who types things one word at a time rather than a model that produces fully formed text. I’m putting in work to sound uncrafted, which I recognise is a bit absurd.
AI theatre
There’s a flip side to all of this that bothers me more, which is people deliberately writing like AI because they think it sounds professional.
You see it in the likes of LinkedIn posts and internal strategy documents. The language is polished, the structure is clean, there are clear headers and bullet points and bold keywords, and the whole thing reads like someone prompted Claude and then copy-pasted the output without changing a word. Maybe they did, maybe they didn’t. The point is that the AI register has become aspirational for some people, a way of signalling that you’re forward-thinking, in trend, and comfortable with the tools.
The AI register is the absence of a voice, dressed up as competence. The people who adopt it are trading whatever personality their writing had for a kind of corporate smoothness that communicates nothing beyond “I, too, have access to a chatbot.” Meh.
What AI didn’t absorb
There’s one style of clear, structured writing that AI almost never defaults to, and it’s the academic argumentative approach. “I argue that…” followed by “First, … Second, … Third, …” with a clear thesis and ordered supporting points. That tradition of prose has centuries behind it, and it actually works well for exactly the kind of structured thinking that AI is supposedly good at.
But AI doesn’t write that way. It reaches for journalistic polish, for marketing cadence, for the kind of inspirational closing that belongs in a TED talk. It writes strictly to impress, producing prose that sounds confident while actively avoiding any actual commitment to a position.
I think this probably reflects what the training data looks like. There’s a lot more journalism, marketing copy, and blog content on the internet than there is academic argumentation, so the model weights skew toward the registers that are most represented. However, I’m speculating here, and I don’t actually know enough about how training works to say this with any confidence. What I do know is that the argumentative style, the one where you state a claim and then defend it with ordered evidence, seems to have been largely passed over, and that feels like a loss.
How I actually use AI to write
Some guides I’ve read suggest feeding Claude your writing samples and letting it draft something in your voice, trusting that style alone will carry the output. That approach skips the thinking and puts all the weight on imitation, which means you end up editing for both substance and voice at the same time. I’d rather front-load the substance so that editing is only about voice.
What I do instead is work through the thinking first. I’ll go back and forth with Claude Code on ideas, structure, key messages, examples, and clarifying questions until I’m happy with the direction. That phase is genuinely collaborative, because the AI is good at surfacing angles I hadn’t considered and pushing me to be specific about what I actually mean.
Then Claude generates a draft based on everything we’ve discussed. And then I go through it sentence by sentence, to rewrite, cut, rearrange, and replace anything that sounds like it came from a model rather than from me. The generated draft serves only as a raw foundation, and most sentences don’t survive contact with my editing pass.
The distinction matters to me because by the time I’m done, the writing is mine. The ideas were shaped collaboratively, but the actual prose went through my hands, which means the voice stays consistent and I don’t have to wonder whether a paragraph sounds off because I missed an AI-ism during a quick skim.
Where this leaves me
Since I lack a clean ending for this, I’ll just say where I’m at.
I’ve given up parts of my writing style that I genuinely liked because they now read as tells for something I didn’t write. That feels like a small but real loss. I’m writing defensively against a tool I also depend on, which is a strange position to be in. And I can see a growing number of people moving in the opposite direction, embracing the AI register as their own, which complicates things because human writing and AI writing are converging from both ends.
I’m still adjusting and I don’t know yet where this settles. I just wanted to write it down while the shift is still fresh enough that I can remember what it felt like before.