You tell it not to. You paste your text into the chatbot, add "don't make this sound like AI," and back comes something that sounds exactly like AI. The reason is worth understanding, because it changes what you do about it. The tells are not choices the machine is making. They are the default shape its sentences take, the way an accent is the shape your mouth makes before you think about it, and you don't fix an accent by asking someone to stop having one.
The instruction reaches the wrong level
When you tell a model "don't write like AI," it agrees, sincerely, and then does it anyway. That isn't defiance. An instruction like that operates at the level of intention, and the tells live below intention, in the reflexes of how a sentence gets built. Ask a person from anywhere to say a word without their accent and watch the same thing happen. They know exactly what you mean, they want to comply, and the sound comes out shaped the way their mouth has always shaped it, because that shaping happens before the part of them that heard your request gets a turn.
A model works the same way. It has read an ocean of text, and certain sentence shapes are simply the grooves it runs in. One is easy to point at. It's the little rhetorical seesaw where a sentence brushes past the plain word for a thing and reaches, in the same breath, for a grander one, waving off "a tool" to land triumphantly on "a partner." That shape feels crisp and confident, which is exactly why it slips out, because it reads like good writing in the moment. The model isn't ignoring your rule when it writes that. The rule is fighting the model's native cadence, and the cadence wins, the same way your accent wins over a single reminder to drop it.
Once you see it this way, "write more carefully" stops being the fix, because careful is not the problem. The shape is the problem, and the shape doesn't respond to sincerity.
Why catching beats preventing
Here is the part that actually helps. Knowing a rule and catching yourself breaking it in the moment are two different skills, and the second one is much harder while you are busy doing everything else.
A writer, human or machine, holding ten decisions at once, what the paragraph is about, who it's for, whether the argument lands, will miss the eleventh, which is the small stylistic reflex nobody was watching for. A reader with one job and fresh eyes will catch it instantly, precisely because catching it is their only job. Talent has little to do with it. Attention is the whole of it, and attention is finite while you are generating and nearly total while you are only checking.
There's real evidence the tells are structural rather than surface-deep. A 2026 study fed AI-written text through a detector, then had the surface cleaned up, the obvious words swapped, the phrasing smoothed. Detection barely moved. What gives the writing away turned out to be the underlying shape, not the vocabulary, which means editing after the fact, on the same pass that wrote it, mostly polishes the tells rather than removing them. The honest conclusion is uncomfortable and freeing at once. You will not instruct your way out of this, so you build a separate step that assumes the writing slipped and goes looking.
The move is a checkpoint, not a better prompt
This is the same lesson as managing any capable tool honestly. Trusting the tool to police itself does not produce good output. Building a step that checks the work does, which is exactly the instinct that scales, whether the checker is a short list you run by eye or a person who reads it cold before it goes out. The owner who understands this stops re-prompting the same paragraph in mounting frustration and starts keeping a checklist of the three tells that always sneak into their newsletter, then reads for those three on the way out the door.
We hold our own writing to this, and we built the enforcement into the machinery rather than the good intentions. Every piece of writing on this site passes through an automatic check before it can publish, one that blocks the piece if it contains the known tells, and we built that gate on the explicit assumption that the writer would fail to catch them all, human or AI. We assumed we would break our own rules, so we made a step that does not depend on us remembering not to, and it catches things weekly. Read back, that sounds like a confession. It isn't. The weekly catch is the design working as intended, and it is the difference between a rule you believe in and a rule that actually holds.
What this changes about using AI to write
None of this means the tool is bad at writing, and none of it means you should stop using it. The draft it hands you is genuinely useful, faster and often better-organized than a blank page. What it means is that "make it sound human" is the wrong instruction to lean on, because it aims at the one level the tool can't reliably reach. Aim somewhere better. Give it your actual voice to work from, the words you really use and the ones you never would, and then, whatever it returns, run the last pass yourself or hand it to someone with fresh eyes and a short list of what to hunt for.
The accent never fully goes away, which is worth knowing before you go looking for the setting that removes it. There isn't one. There is only the habit of listening for it on the way out, and that habit is learnable in an afternoon, which is more than you can say for losing an accent you were raised with.