Let's get the disclosure out of the way: I wrote this with AI. This exact post went through the three rounds in the title, so if it reads clean, that's the point. If it doesn't, that's the more interesting comment.
It started some paragraphs, cleaned up others, and helped turn a pile of dictated notes into something readable. If you find a sentence that sounds suspiciously polished, tell me where. I'm curious which ones I missed.
I don't really think of an article as a finished object anymore. I write one version, then it has to travel. LinkedIn needs a different opening. X needs half the words. A newsletter can keep the context that a social post has to lose. Sometimes the same idea ends up in five places, each with its own limits and conventions.
This is where AI saves me a lot of time. It can cut, rearrange, reformat, and produce ten versions without getting bored. The original writing is another matter. I still don't trust it to do that alone.
Writing is fashionable to declare dead right now. Why bother learning when AI can generate a post in seconds? It reminds me of the way people talked about SEO around 2016. SEO didn't disappear. It changed, picked up a few new abbreviations, and carried on taking up everyone's time.
I suspect writing will do the same.
AI-generated text is already recognizable, although it isn't always easy to explain why. Certain phrases give it away immediately. "Here's what I found." "Not generic. Not robotic. You." "Honestly," placed at the start of a sentence as though everything before it had been dishonest.
The model reaches for these phrases when it needs a bridge but has nothing specific to add. Once you've edited enough generated text, you start spotting them before you've properly read the sentence.
The structure is often more revealing than the vocabulary. A text begins in ordinary paragraphs, then gradually turns vertical.
Like this.
One sentence per line.
Each thought given the same weight as the last.
Looks clean, doesn't it. Reads with a certain rhythm.
Then the paragraph snaps back to full width and you notice you'd stopped following an argument and started being walked down a list, with every thought given the same space and emphasis whether it earned it or not. It reads especially well on a phone, which is part of the problem.
You can tell the model to write in paragraphs, vary the sentence length, and avoid staccato. It will follow the instruction for a while, then drift back. That's the style it has seen rewarded across a huge amount of online writing.
The current panic about AI writing has also made careful human writing look suspicious.
I once wrote a comment entirely by hand. It had proper paragraphs, a clear structure, and punctuation exactly where I wanted it. An automated moderator removed it almost immediately.
No person read it and decided it was bad. A system decided, from the pattern alone, that it looked generated.
The comment was too orderly. The paragraphs were balanced, the argument moved cleanly from one point to the next, and apparently a real person was not expected to write that smoothly. I had been classified as a robot for taking the time to edit.
Since then, I've been skeptical whenever someone presents a list of definitive "AI tells." Sometimes those patterns do come from AI. Sometimes they come from people who know how to construct a paragraph.
The em dash is the most ridiculous example. Apparently it belongs to language models now. This thing, the one I just used, is supposed to be a confession.
I've been typing it for fourteen years. My fingers learned Shift-Option-hyphen long before I knew what a large language model was. I'm not going to retrain them because automated detectors have developed an anxiety about punctuation.
The dash is a poor test anyway. AI usually reveals itself at a larger scale: a rhythm that never changes, groups of three arranged a little too neatly, repeated "not X, but Y" constructions, and sentences that sound complete without containing much of anything. A single punctuation mark tells you very little.
So what do I actually do when AI has touched a draft?
First, I read the whole thing aloud. My eyes will accept a sentence because it looks polished. My ears are less forgiving. They catch the places where the language is doing more work than the idea.
Then I check the meaning sentence by sentence. What is this claiming? Where did that claim come from? Is it saying something concrete, or is it only moving me toward the next paragraph?
Unsourced facts either get a source or disappear. Tenses and numbers get checked separately because models mix both up with remarkable confidence.
After that, I give the draft to a second AI with a simple instruction: find and remove anything that sounds generated. It usually spots phrases I no longer notice because I've been staring at the text for too long.
Then I edit its edit.
In removing clichés, the second pass often removes meaning, tone, or some awkward detail that made the paragraph sound like me. It tends to solve messiness by making everything neutral. I compare the versions, put back what mattered, and rewrite the rest by hand.
That final pass is usually where the text stops feeling generated.
I have a personal list of things that make me pause: "delve," "seamless," "game-changer," empty bridges such as "here's what I found," too many mirrored contrasts, and staccato that runs through a whole piece. None of these is automatically forbidden. People use them too. But each one makes me check whether the sentence exists because I meant it or because the model needed something plausible in that spot.
The process isn't as new as it sounds. When I was learning to write, I did roughly the same thing with human editors.
Several of my friends are editors, and they were not gentle. They crossed out clever sentences, asked where facts came from, and pointed to the exact line where I had stopped saying anything. They argued with me when I defended a phrase only because I liked how it sounded.
Back then, I was training myself through repeated feedback. Now I'm training a model on top of that. The method has barely changed: show examples, explain what failed, revise, and don't accept the first pass just because it's grammatically correct.
I don't ask AI to "write in an interesting voice." That instruction usually produces a polished average of everything online. Instead, I give it concrete examples of writing whose rhythm I like and explain what I want it to notice: how long the paragraphs run, where the author gets specific, how jokes are placed, and which ideas are allowed to stay slightly rough.
Even then, the best results rarely begin with a blank page.
AI works much better when I give it a raw draft: dictated, repetitive, badly ordered, but mine. It can untangle the structure and cut the repetition without having to invent the material. Ask it to start from nothing and you generally get something competent that you'd never have chosen to say yourself.
The human still has to bring the observation, the story, and the reason for writing at all. A model is good at combing through that material. Give it a blank page and ask it to find the point, and that's where it starts to fall apart.
Once the article is finished, the balance changes. It still needs to become a LinkedIn post, a shorter version for X, an email, maybe a few replies. That work is repetitive, rule-based, and easy to verify. I hand it to AI without regret.
Publishing is repetitive too. I handle that through Publora now, which I'm part of, so weigh the mention accordingly. But the distinction holds no matter the tool: writing an idea and reproducing it across formats are two different jobs, and only one of them is the reason I started.
So no, I don't think writing is dead. I think the work around it is changing. There's more generated text to edit, more polished emptiness to catch, and more reason to know what your own voice sounds like before you ask a model to imitate it.
It's slower than typing "write me a post" and publishing what comes back. It also gives me something I can still recognize as mine.
AI helped write this article. Which sentence gave it away?
Top comments (12)
I think the most interesting risk isn’t that AI sounds like us—it’s that, over time, we start sounding like AI. Once you’ve edited hundreds of AI-assisted drafts, it’s surprisingly easy to internalize the same rhythms, transitions, and sentence structures yourself. The challenge isn’t only preserving your voice today, but noticing when your own writing gradually converges toward the model’s statistical average.
Ooooh, a hundred percent yes! This works perfectly if you're running not a personal page but, say, a company's — when the tone of voice, structure, and rules are already established. But personal ones should still stay personal. Plus 1000 to your comment.
Exactly. A company is trying to maintain a consistent brand voice, while a personal account is trying to preserve an individual one. The irony is that the better AI gets at imitating us, the more valuable our genuinely human quirks become. Those little imperfections are often what make a piece memorable.
100%!
Absolutely. As I had the AI continuously revise and iterate, I found that its writing started to resemble my style, and my writing began to resemble its style more and more as well. What’s even more disappointing is that one day my friend showed me her article, but I didn’t recognize that it was text that had been run through an AI model several times.
Oh yeah, it's getting harder and harder to create texts that are unique in style. And even if you manage to, someone will still say it's definitely made by AI.
You can tell right away when certain phrases are used I'm the one behind trimy.io and I can say I spend a lot of time rewriting stuff to make it sound less like it was written by a machine before I put my name on i, there's just something about the way the words are put together the rhythm is too perfect, the lines are too neat. I can spot it in just a few seconds now.
Reading it out loud is the best test. If it sounds like something I'd say to a person, it stays. If it sounds like a summary of what I'd say, it goes.
Great! We're alike in this — I do the same :)
I couldn't agree more, Eugeniya . I've had a version of this happen to me too, an essay about AI detection getting flagged by DEV.to's own moderation for the same reason. The "too orderly" flag is the same failure from the opposite direction.
The detector isn't scanning for AI, it's scanning for structural cleanliness and treating that as a proxy. Works passably on AI because AI is trained on edited prose. Fails hard on you because edited human prose looks the same to the model.
Did the mod ever say which feature tripped it — paragraph balance, the em dash, something else or just "reads generated," no breakdown?
Exactly the problem. There's no breakdown at all. The filter just says it doesn't pass and leaves it at that. No explanation, no feature it didn't like, nothing to push back on.
My guess is it's the formatting you mentioned — the neat paragraphs, maybe the em dashes — but that's all it is: a guess. The filter never tells you what it flagged. That's the frustrating part. You can't appeal a decision when you don't know what the decision was based on.
The only other option is opening a thread with the human moderators. That takes time, and there's no guarantee it'll change anything.
So for now I'm just sitting here, with a comment that won't publish because a bot decided it looked wrong.
The fact that it won't even name the feature is worse than my case. A wrong verdict you can see the reasoning for is at least falsifiable, something you can actually argue against. A wrong verdict with no reasoning attached isn't a verdict, it's just a wall.
Does the moderator thread actually surface the flag, or does it stay a black box all the way up?
Unfortunately, it stays a black box :(