TL;DR: Removing the AI tells is an engineering problem. It can be broken into four stages plus one automated lint script, and it took me three months to turn it into something repeatable. But tools can only get you to “doesn’t sound like a machine.” They cannot get you to “sounds like you.” The hardest thing to quit turned out to be antithesis: a pattern that was mine to begin with, not one of the cheap AI tics.

“A man’s style is his own portrait.”

— Wittgenstein, Culture and Value

I was staring at the thirty-first em-dash on my screen, and something felt wrong.

The article was about how automation can turn around and constrain the people who depend on it. I was satisfied with the content; the argument held. The em-dashes were the problem. One appeared every few lines, used to pivot, to pause, to finish a thought I hadn’t quite closed. Each one looked reasonable on its own. Put together, the whole piece read like a machine had written it.

So I changed them one by one. Where a break was possible, I used a period. Where something needed adding, I rewrote it as a complete sentence. Where a pause was genuinely necessary, I kept the punctuation. Thirty-one em-dashes, none remaining. By the time I finished that night, something had come clear: the AI tells aren’t about single words. They are more like a set of habits: subtle at first, and by the time you notice, already soaked through the whole piece.

After that, I spent three months turning “how do I keep my writing from sounding like AI wrote it” from a gut-level editing instinct into a process I could run again and again. That process, including the times it pushed back, is what this piece is about.

But I need to say upfront: this method only does half the job.

It can edit a piece until it no longer sounds like a machine. It cannot automatically edit it until it sounds like you.

Think of cosmetic surgery. If you follow the golden ratio precisely, you can bring a face to a certain standard of attractiveness. But the far end of that standard is often a kind of convergence, because everyone is following the same ratio. Writing works the same way. Use the same set of de-AI tools, built on roughly the same underlying models, and you tend to arrive at the same kind of clean. Rhetoric, tone, argumentative structure, even style start to rhyme across writers. The prose becomes readable; it just becomes hard to remember whose it was.

So this series splits in two.

The engineering half covers how to reduce the AI tells. I’ll be specific enough that you can borrow the approach and plug it directly into your workflow.

The other half is about what comes after the reduction: how writing grows back into something distinctly yours. That part is slower. It requires practice. There is no shortcut.

Spotting the AI Tells Is Easy. Quitting Them Is Not.

Articles that teach you to recognize AI-flavored writing are abundant. Red-flag words, em-dashes, negation parallelism, fake-sincerity openings: every pattern has been catalogued somewhere. These lists are useful. They all share the same vantage point, though: the reader’s, or the detector’s. You hold the list and examine someone else’s work.

Quitting it in your own writing is a different problem.

You are hunting your own habits, not someone else’s. And AI-assisted writing gradually learns your voice. When what it produces happens to match the sentence patterns you naturally use, the checklist stops working. Your eye passes right over that line.

It took me a while to understand: the hardest AI traces to eliminate are the ones that look most like me, not the ones that look least like me.

Stage One: Scanning by Eye, on Instinct

The first method was blunt. After finishing a draft, I read through it looking for familiar AI patterns and rewrote what I found. I started with a list of six: binary antithesis, escalating triplets, empty subjects, generic filler phrases, false-vulnerability openings, ceremonial closings. Before anything went live, I ran through that list once.

This held for a while. Em-dashes came later. At the time, a lot of AI-generated writing was dense with them, and my first draft of that piece had thirty-one. I had scanned through several times before I caught it. I saw them; I just didn’t register them as a problem at first.

The list grew from six items to seven. But I knew the eye always misses something.

Stage Two: People Miss Things, Especially What Looks Most Like Themselves

Eventually I ran a full audit. I picked ten published articles and worked through the checklist item by item, counting.

The results were interesting. The cheaper AI tells (empty subjects, generic fillers, false-vulnerability openings, ceremonial closings) came up zero across all ten. Those were long gone. Em-dashes were under control too; newer pieces all scored zero, with only earlier writing still running over.

The one pattern I couldn’t quit was binary antithesis: the “not X, but Y” construction. It still appeared in my most recent work, with one piece hitting seven instances.

Why that one specifically? Honestly, I can’t blame AI entirely. Antithesis has always been my habit. My thinking tends to move by structural contrast. Frames like “this isn’t a personal failure, it’s a systems problem” show up in my arguments regularly. Antithesis is just the most natural container for that kind of thinking.

One antithesis is a blade. Seven is a verbal tic. A tic dulls the blade. The one carrying your central argument is worth keeping. The other six are mostly habit. They should become questions, plain statements, or anything else that breaks the mold.

The audit made one thing clear: anything a script can catch reliably, I should stop pretending my eyes can handle. I wrote a script to cover the five items that admit precise detection: antithesis, em-dashes, runs of enumeration commas, runs of short sentences, and filler phrases. It runs automatically before any piece goes out. That script, together with the manual checklist, gradually became a writing skill I could actually rely on.

Machines are good at counting. People are good at judging. Hand the counting to the machine. Keep the judgment.

Stage Three: Not Just Removing the AI Tells, But Growing Back the Human One

Once the system was running, I noticed I had been doing subtraction the entire time. Find an AI trace, cut it. Find another, cut that. Subtraction has a floor, though. A piece can become clean enough that almost no AI traces remain and still feel cold.

I had misunderstood what it means to refine something. Refining is not minimizing word count. The sense of flab in a piece usually has nothing to do with length. It comes from words that aren’t doing any work. A metaphor, a passing observation, a physical detail: if it is carrying something, whether rhythm, presence, or warmth, it earns its place. What looks like cutting deadweight sometimes turns out to be cutting body heat.

Once writing is clean enough, what it lacks is temperature.

My default mode is expository: breaking a position into a sequence of arguments. That is a strength, but it has a cost: exposition runs cold by nature.

What warms writing is testimony, not more reasoning. Bringing an experience back as a scene with setting, body, and movement. Readers respond to something happening; they struggle to feel much about a string of propositions.

This is not a call to turn every piece into an essay. What I do now is keep the argument as the skeleton, but find a few passages in each piece where something that actually happened can come in and give the argument a little heat.

I wrote a piece on worldview that was argument all the way through. Looking back, the part that had warmth was the mention of a monthly utility bill. That was where readers found something to hold. People remember the moment when an idea touches their life, rarely the argument itself.

That was when I understood what self-checking is actually for. The goal is not to delete words. It is to turn the forms that run automatically back into forms I have chosen.

Stage Four: Lock It In So Memory Isn’t Required

The last step was making sure none of this would fade over time or break down when my writing context changed.

I consolidated all the writing-quality rules into one place, a single source of record. Long-form pieces, social posts, other formats: anything involving writing quality points back there. No scattered copies. I also filled in a gap I had consistently missed: sentence completion. The AI-tells checklist handles one question: does this sound like a machine? The sentence-completion work handles another: is this sentence finished, and is the meaning precise? I organized it into ten categories: restoring dropped subjects, closing open verbs, returning each punctuation mark to its proper role, converting bloated nominalizations back into verbs. These are the moves I make most often when editing my own drafts.

Later I brought translation into the same framework. A Chinese draft with the antithesis removed will still produce “not X, but Y” in the English version. That is a separate layer of linguistic habit, invisible to the Chinese original. So I attached a negative-pattern list to the English and Japanese translation passes, carried automatically, so those constructions don’t quietly revive themselves in another language.

Removing the AI Tells Is Engineering. It Is Also Craft.

Looking back, the biggest gain from those three months was a clearer understanding of what writing actually is, not an additional rule to follow.

There is no point at which removing the AI tells is finished. Quit em-dashes, and antithesis remains. Suppress antithesis, and it resurfaces in translation. Models improve, and a new set of tics arrives.

Writing is both a process with version control and automated checks, and a craft practiced by hand, because at its core, writing is practicing thought. It is a conversation with a reader. It is thinking becoming habitable.

When I write, I ask myself: what am I trying to signal? Who am I talking to? In the immensity of what has already been said, what can I add? AI can generate opinions quickly. How I present what I actually think is still my decision.

The thirty-one em-dashes in that early piece I changed by hand, one at a time. That kind of work now runs in batch through a script. When I see an em-dash today, I’m no longer asking whether it sounds like AI. I’m asking: if I take it out, does what remains sound more like me? Or have I been using AI to impersonate someone I’m not?