The sentence habits that make a post read as machine written

The AI smell has a mechanical cause you can edit

What readers call the AI voice is a small set of sentence habits, and all of them are visible on the page before you get anywhere near vocabulary. That distinction is the whole of the practical problem. A smell cannot be edited and a habit can, so the useful move is to stop describing the impression and start naming the line that produced it.

The complaint threads on this subject are accurate about the symptom and useless at the desk. They tell you the post had invented suffering, a sudden reversal and a tidy lesson, which is true of a great many posts and never tells you which sentence to change. The humanizer tools have the opposite problem, because they act on the draft without naming anything, so nothing is learned and the next draft arrives with the same habits.

The six habits below share one cause. Each of them lets a sentence sound finished without committing to a relationship between two ideas. A fragment stack never has to say why the second thing followed the first. A group of three never has to say which of the three matters. A closing moral never has to say who should now do what. Prose that avoids all three commitments reads as competent and says very little, which is precisely the impression readers are describing when they say something feels machine made.

Vocabulary is the last place to look. Word choice is the easiest thing for a model to vary and the easiest thing for a tool to swap, which is why every humanizer works there. Rhythm is what a reader actually recognises, and rhythm survives any number of synonyms.

The Personeur Tell Sheet: six tells and the repair for each

Six habits account for most of what readers flag as machine written, and each has a repair that does not require rewriting the argument. Work down the list in the order given, because the first two are audible from across a room and the last two are the ones that survive a careless edit.

The sheet is diagnostic rather than prohibitive. Every one of these devices is legitimate once in a while, and a writer who has banned them all is producing a different kind of flat prose. What makes them tells is density. Two of them in one post is a style. Five of them in one post is a machine, or a person imitating one.

Tell one: dissolving a fragment stack back into speech

The repair for a fragment stack is to restore the connective it removed. The stack is not really a style choice, it is a way of avoiding the sentence that explains why one thing led to another. Rejoin the fragments with because, so, until or which meant, then read what you are left with. If the joined sentence sounds obvious, the stack was concealing the fact that it was obvious.

It is worth knowing why the habit is everywhere. The cadence is cheap to produce and it feels decisive, which is why advertising has used it for decades, and it arrives by default in drafts written by a model because that same advertising is heavily represented in what those models have read. Neither origin matters at the desk. The repair is identical whether a person or a machine put the beats there.

Three stacks dissolved. The examples are composites written for this article, not client work.

Look at the third column rather than the second. Every repair drags one new specific into the post, because a connective needs something real on both sides of it. That is the actual benefit of this pass. The prose stops sounding machine made as a side effect of becoming more particular, which is a better outcome than sounding human by decoration.

Tell two: the one-line paragraph, and the ratio that keeps it working

A standalone line is a genuine emphasis device and it stops being one when every line is standalone, because emphasis is a ratio and not a format. Keep roughly one standalone line per hundred words of post and the device holds its meaning. Go past that and the white space is telling the reader that everything is important, which reads the same as nothing being important.

Work it out on a normal post rather than taking the ratio on trust. Assume 220 words, which is a reasonable length for a text post meant to be read rather than glanced at. Eight standalone lines in that post is an emphasis marker every 27 words, or roughly every second sentence. Two standalone lines in the same post is one marker every 110 words, which a reader still feels as a change of pace. Those are assumptions about your own post, not measurements of anybody else's.

The same 220-word post, at four different densities of white space.

There is a real tension here and it deserves stating plainly rather than resolving by slogan. A feed rewards white space, because a reader scanning between meetings needs somewhere for the eye to rest, so the advice to break lines has a genuine basis. The ear punishes the same habit. The resolution is that white space should mark the two or three lines you want a skimming reader to stop on, which makes it a decision about what matters rather than a default setting. That is the same decision described from the reader's side in how a busy reader actually scans a post.

Three more tells that share one cause: the dash, the triple, the reversal

These three come from a sentence trying to sound balanced instead of trying to say something, and all three are repaired the same way. Remove the balancing device, then check whether the sentence still contains a claim. Often it does not, and the tell was holding the place where the thinking should have been.

The long dash

The em dash is not an error and it is overused by almost everybody writing quickly, which makes it a signal rather than a fault. Three appearances in a short post is the point at which a reader starts noticing punctuation instead of the argument. The repair is to ask what job the dash was doing, then use the mark that does that job properly.

Four jobs the long dash gets used for, and the mark that actually does each one.

The rule of three

Count the items and then name the weakest one out loud. If the third item exists because two sounded thin, it is rhythm and it goes. If all three are genuinely load bearing, add the fourth, because a real list of three in a post containing no other triples reads as a list, while a triple in a post full of triples reads as a tic. The test is density again rather than the device itself.

The false reversal

A post that opens by announcing the problem was never the tooling, it was the process, has told you nothing at all unless somebody in the room was actually blaming the tooling. If nobody was, the first half exists only so the second half arrives sounding earned. Delete the denial. What survives is either a claim you can defend or an empty line, and finding out which one it is takes about four seconds.

Humanizer buttons and AI detectors both fail, for different reasons

A humanizer tool changes vocabulary while every tell on the sheet is structural, so the tells survive the pass untouched and the word choice gets slightly stranger. A detector scores how predictable each next word is, which punishes plain writing and rewards writing that wanders, so it cannot answer the question you are asking it.

Watch what a humanizer does to a fragment stack. It swaps one adjective, shortens a phrase, perhaps moves a comma. The three beats are still three beats and the standalone lines are still standalone. The colleague who told you the post felt machine made was responding to the rhythm, and the rhythm has not been touched by any of that.

The detectors have a deeper problem. They measure predictability, so text where each next word is easy to guess scores as machine written and text that meanders scores as human. An executive who writes plainly, repeats their own terminology and uses the same three words for the same idea across a year of posts is predictable by design. That is what a consistent voice is. Running a consistent voice through a detector produces a number about clarity and not about authorship.

Readers notice something different again: repetition of shape across posts, an absence of anything a competitor would dispute, and prose that could have been written about any company in the sector. None of that is a vocabulary problem either. Whether to tell an audience that a writer is involved is a question of policy rather than craft, settled once in a line you can live with, which is the subject of whether to disclose a ghostwriter.

A worked rewrite: one paragraph, three passes

Here is the whole method applied to one paragraph. It is a composite written for this article rather than client work, and it carries five of the six tells on purpose, which is roughly the density that makes a reader say a post feels machine made without being able to say why.

Last quarter we lost a deal we should have won. It stung. But it taught me something about how we sell. We were talking about features, not outcomes. The customer wanted certainty, not a demo. Three things changed after that call, and every one of them mattered.

Pass one, dissolve the fragments and the false reversals

The stack sits in the second and third sentences, and rejoining them forces the paragraph to say what was learned rather than announcing that learning occurred. The two false reversals are the features against outcomes line and the certainty against a demo line. Both deny a position that nobody in the story held, so both denials come out and the claims stay.

Last quarter we lost a deal we should have won, and the loss taught me that we were selling features while the buyer was trying to reduce her own risk.

Pass two, put the specifics back where the tells were

Every tell you remove leaves a gap, and the gap is exactly where a specific belongs. Which deal, how far it got, what the buyer actually said, and who she said it to. This is the pass that decides whether the post was worth writing, because a paragraph with no recoverable specifics was never about anything.

We lost a deal in the final review last quarter after four months of work. The buyer told our sales lead she could not defend the choice internally if the rollout slipped, and we had spent the whole cycle demonstrating features.

Pass three, cut the moral and end on the detail

The closing line of the original promised three changes and delivered none of them, which is the tidy moral doing its usual work of sounding like an ending. Cut it and end on the specific a reader could repeat to a colleague, or on the question you have genuinely not resolved. An unresolved question also gives the comments something to do besides applauding.

We lost a deal in the final review last quarter after four months of work. The buyer told our sales lead she could not defend the choice internally if the rollout slipped, and we had spent the whole cycle demonstrating features instead of showing her how to defend it. I still do not know whether we could have found that out in week one, or whether it only becomes sayable in the final review.

The paragraph went from 47 words to 74. Every added word is a specific: a timeline, a named objection, the person who raised it, and one question left open. Not one repair required a better vocabulary, and the finished version is easier to say out loud than the original was.

The read-aloud pass, and the four hours a year it costs

Reading the draft aloud finds more of these tells in ninety seconds than any tool will, because every one of them is audible before it is visible. A fragment stack sounds like an announcer. Standalone lines sound like a countdown. A tidy moral sounds like the last line of an advertisement, which is where most of us learned the shape in the first place.

The cost is small enough to check with arithmetic rather than argue about. Assume a 220-word post and an unhurried reading speed of about 150 words a minute, which puts one post at roughly 90 seconds. Three posts a week is four and a half minutes, and 52 weeks of that is under four hours a year. Assume the rewriting the pass triggers takes about three times as long as the reading, and the whole discipline costs something like sixteen hours a year. Those are stated assumptions rather than measurements, so replace the reading speed and the post length with your own and rerun it.

One decision sits upstream of all of this. A tell repaired inside the wrong format is still the wrong post, because a 900-word argument compressed into six standalone lines will read as machine made no matter how carefully each line is repaired. The shape and the length are worth settling before the first sentence, which is the subject of choosing the length before you write.