AI writing tells: the sentence level

This is part two of a five-part series on AI writing tells, working down from the surface to the spine. Part one took the word level and argued that scrubbing the vocabulary is the shallowest possible fix. This piece assumes you have already done that scrub. The “delve” is gone, the em dashes are thinned, the inflated adjectives are cut. And the prose still reads like a machine wrote it. This is why.

The rhythm gives it away

Read a few paragraphs of clean-vocabulary AI prose out loud. Not skimmed, read aloud, the way you would read to a child. Within a minute you will hear it: a rhythm that is too even, a cadence that never surprises you, sentence after sentence landing at the same length and the same weight with the same little internal shape. The words are fine. The music is wrong.

That is the sentence layer. It is the first place the machine gives itself away after the words are clean, and it is harder to fix because there is no list to run. You cannot find-and-replace a rhythm. You have to hear it, and then you have to know what specific constructions are producing it, because the bad music is not vague. It comes from a handful of specific sentence-level habits, and once you can name them you can hunt them.

This article names them. Each one is a construction the machine reaches for constantly, and each one, repeated, flattens the rhythm into that telltale evenness. Learn the constructions and you can hear the prose the way an editor does, which is the only way to fix it.

The “not just X but Y” reflex

Start with the most recognizable one. “It’s not just a tool, it’s a way of thinking.” “This isn’t merely about profit, it’s about purpose.” “She wasn’t just tired, she was exhausted.” The machine loves this construction with a depth that is hard to overstate. It is the single most common sentence shape in AI prose after the plain declarative.

The reason it loves it is revealing. The construction lets the machine sound like it is making a sophisticated distinction, a correction, a deepening, without actually having one. “Not just X but Y” promises that Y is a meaningful escalation past X, and most of the time it is not. “Not just tired but exhausted” is not a real distinction, it is the same idea said twice with a fake pivot in the middle. The construction creates the shape of insight around a place where there is no insight.

One of these in a piece is fine. Humans use it. The tell is the frequency, because the machine cannot stop reaching for it, and after the fourth or fifth “not just, but” the reader’s ear starts to register the pattern even if the reader cannot name it. The fix is not to ban the construction. The fix is to find every instance and ask whether the pivot is real. If Y is genuinely a leap past X, keep it. If Y is just X again in a bigger word, cut the whole thing down to a single plain statement, because the construction was building a staircase to a floor that was already there.

The rule of three

The machine counts to three. It does it everywhere, and once you see it you cannot unsee it. Lists arrive in threes. “Clear, concise, and compelling.” “It was cold, dark, and unwelcoming.” Examples come in threes. Adjectives stack in threes. Whole paragraphs build to a closing triplet, three short phrases in a row, because the machine has internalized that three is the rhythm of rhetorical completeness.

Three is a real rhetorical device. Good writers use it on purpose, at a moment that earns the drumbeat, and the deliberateness is exactly what makes it land. The machine uses it as a default, the way it uses the em dash, three items where two would do or four would be truer, because three feels balanced and the machine reaches for balance over truth every time.

Here is the test. Count the items in a list and ask why there are that many. If a human wrote it, the count follows the truth, two things because there were two things, five because there were five. If the machine wrote it, the count is almost always three, regardless of how many real items exist, because three is the shape and the shape came first. The fix is to let the real number govern. Some lists are two. Some are six. When you stop rounding everything to three, the prose immediately sounds less composed and more like a person actually counting real things, which is the goal.

Hedges and throat-clearing

The machine cannot start a thought cleanly. It clears its throat first. “It’s worth noting that.” “It’s important to understand that.” “One might argue that.” “In many ways.” “To a certain extent.” These openers add nothing and the machine deploys them constantly, because a bare declarative feels abrupt to it and the hedge softens the entry.

The damage is double. First, the hedges flatten the rhythm, because they are all the same shape, a soft windup before the actual sentence. Second, they leak the machine’s core problem, which is that it does not trust its own statements enough to make them flat. “It’s worth noting that the policy failed” is a machine sentence. “The policy failed” is a human one. The human writer states the thing. The machine asks permission to state the thing.

Cut every hedge. This is one of the few sentence-level fixes that is almost mechanical, because the hedges genuinely add nothing and the sentence is always stronger without them. “It’s important to understand that X” is always just “X.” The only thing the hedge ever did was lower the confidence and flatten the music. When you strip them all out, two things happen at once: the rhythm sharpens, because the soft windups are gone, and the voice gains confidence, because every statement now stands on its own. Throat-clearing is the easiest deep tell to fix and one of the highest-impact, which is a rare combination.

The gerund opening

The machine loves to open a sentence with an -ing word doing abstract work. “Understanding the market requires patience.” “Building trust takes time.” “Navigating these challenges means staying flexible.” The gerund-as-abstract-subject is a deep machine habit, and it produces a particular kind of limp, front-loaded sentence that never quite gets moving.

The problem is that the gerund opening defers the actual subject. “Understanding the market requires patience” is really just “the market is hard to understand” or “you need patience to understand the market,” but the machine fronts the abstraction because it sounds more elevated. The result is a sentence that starts in a fog of process language before it tells you what it is actually about.

The fix is to find the real subject and put it first. “Understanding the market requires patience” becomes “the market takes patience to read.” “Building a team requires trust” becomes “a team runs on trust.” The moment you demote the gerund and promote the concrete noun, the sentence stands up. Do this across a whole piece and the front-loaded fog clears, and the prose stops sounding like a management seminar.

The “which” chain and the run-on

The machine builds long sentences by chaining clauses, and its favorite link is “which.” “He took the job, which paid well, which meant they could finally move, which was what she had wanted all along.” Each “which” hangs another car on the train, and the sentence rolls on past the point where it should have stopped. The same thing happens with “and,” with “but,” with “so,” with “that,” the machine coupling clause after clause because it does not have a strong instinct for where a sentence ends.

Humans write long sentences too. The difference is that a good long sentence has architecture, a shape that builds and resolves, where the length is doing something. The machine’s long sentence has no architecture. It is just accumulation, one clause after another until the thought runs out, and the reader arrives at the end having lost the beginning.

The fix is brutal and simple: break the chain at every link that can stand alone. “He took the job. It paid well, which meant they could finally move. That was what she had wanted all along.” Three sentences where there was one, and now each idea gets its own weight and the rhythm has variation, short against long, instead of one endless roll. The “which” chain is a particularly good thing to hunt because it is easy to find, search the piece for sentences with two or more “which” clauses, and every one of them is a run-on in disguise.

The “and” version of this is worse, and at length it becomes the machine’s single most pervasive sentence habit. Left unconstrained, the machine writes in clauses chained by “and,” and the chain does not stop, and every clause gets the same weight as the one before it, and the result is a breathless undifferentiated rush where nothing is subordinated and nothing is emphasized and the reader is pulled along through a flat sequence of equally-weighted facts. Read that previous sentence again and you can feel it. That is the machine’s natural rhythm when nothing tells it otherwise, and it dominates an unconstrained draft to a degree that has to be seen to be believed, paragraph after paragraph of and-and-and, because the machine generates forward, clause by clause, with no instinct for shaping, and “and” is the cheapest possible link between one clause and the next. A human varies. Short sentence. Then a longer one that builds and turns. Then a fragment. The machine, on its own, does none of that, and the uniform and-chain is the prose equivalent of a voice that never changes pitch.

Parallelism run into the ground

The machine loves parallel structure, and like everything it loves, it overuses it until the love becomes a tell. Parallelism is when you repeat a grammatical shape across phrases or clauses for rhythm and emphasis: “We came, we saw, we conquered.” Used at the right moment, it is one of the most powerful tools in prose, because the repeated shape builds a drumbeat the reader feels. The machine deploys it constantly, in places that have not earned the drumbeat, until the prose starts to feel like it is perpetually building to a climax it never reaches.

You see it in sentences like “It is about the work you put in, the risks you take, and the person you become.” Three parallel phrases, each “the X you Y,” marching in formation. One such sentence is fine. The machine produces them in clusters, paragraph after paragraph reaching for the balanced parallel triplet, and the cumulative effect is a kind of oratorical sameness, as if every sentence were the closing line of a graduation speech. Real prose mixes parallel and non-parallel constructions, using the parallel one rarely, at the moment it counts, so that when the drumbeat comes the reader feels it. The machine beats the drum continuously, and a drum beaten continuously stops being a drum and becomes a hum.

The fix is to find the parallel constructions and keep only the ones that earn the emphasis. Most can be broken into ordinary, varied sentences that carry the same content without the relentless balance. Save the parallel structure for the one moment in the piece where you actually want the reader’s pulse to pick up, and it will work, because you will not have spent it everywhere else.

The sentence that is all the same length

Here is a tell that hides in plain sight: the machine writes sentences of remarkably uniform length. Not identical, but clustered in a narrow band, most of them medium, fifteen to twenty-five words, with few very short ones and few very long ones. This uniformity is a big part of the flat rhythm, and it is invisible if you are reading for meaning rather than for length.

Human prose varies sentence length hard, and the variation is where the rhythm lives. A long, winding, accumulating sentence that builds and builds, followed by three words. Then a medium one. Then two words. The short sentence after a long one hits like a slap, and it only works because of the contrast. The machine almost never gives you that slap, because it almost never writes the three-word sentence, because a three-word sentence feels incomplete to it and it wants every sentence to be a proper, fully developed thought. So it stays in the medium band, and the prose never speeds up or slams to a stop, it just rolls along at one steady pace.

The fix is to introduce short sentences on purpose, especially after long ones, and especially at moments of emphasis. Find your most important point and put it in the shortest sentence in the paragraph. Find a long machine sentence and break a piece of it off into a fragment of a few words that lands hard. The very short sentence is one of the most powerful tools in prose and one the machine systematically avoids, so reaching for it is both a fix and a tell in reverse, a thing the human does that the machine will not. Vary the lengths violently and the rhythm comes alive, because rhythm is contrast, and the machine’s even band has none.

The weak word at the end

This is the most subtle sentence-level tell and the one that separates people who have an ear from people who do not. The last word of a sentence carries weight. It is the word that rings after the sentence ends, the one the reader’s mind rests on for a beat before moving to the next. Strong writers know this and engineer their sentences so the last word is doing work. The machine does not know this, and it ends sentences on function words constantly.

“That was the thing I had been worried about.” The sentence dies on “about,” a preposition, the weakest possible landing. “I didn’t really know what any of it was for.” Dies on “for.” “It was more complicated than it seemed at first.” Dies on a limp adverbial phrase. Every one of these could end on a word that rings, and the machine ends them on a word that thuds, because it builds sentences front to back without ever feeling the weight of the final beat.

The fix takes an ear, which is why it is the last one in this article. You go through and read for the last word of each sentence, and wherever it is a preposition or a filler verb or a word that just trails off, you rebuild the sentence so it lands on something solid. “That was the thing I had been worried about” becomes “that was the thing I feared.” Now it lands on “feared,” a word with weight. Do this across a piece and the prose gains a quality that is hard to describe but easy to feel: each sentence closes with a small click instead of a fade, and the cumulative effect is writing that sounds intentional instead of generated.

The sentence that exists to set up the next one

The machine writes a lot of sentences whose only job is to introduce the sentence after them. “There is one thing you need to understand first.” “Let me explain what I mean.” “This raises an important question.” “Consider the following.” These are not sentences with content. They are runways, built so the next sentence can take off, and the machine builds them constantly because it learned that prose should ease the reader from point to point.

The tell is that you can delete the runway and lose nothing. “There is one thing you need to understand first. The market does not care about your effort.” Cut the first sentence and you have “The market does not care about your effort,” which is stronger, because it lands without warning instead of being announced. The runway sentence lowered the impact of the real sentence by telling the reader to brace for it, and bracing is the enemy of impact. The best sentences arrive without a runway, surprising the reader a little, and the machine cannot resist building the runway because it does not trust the reader to handle an unannounced landing.

Hunt these by looking for any sentence that points forward instead of saying something. “Consider the following,” “here is the key insight,” “let me be clear,” all of them are pointing at content that has not arrived yet rather than delivering content now. Cut them, and let the sentence they were introducing stand on its own. The piece gets faster and harder, because you have removed a layer of throat-clearing between the reader and the actual material, and the actual material is what they came for.

Why the sentence layer matters more than the word layer

Here is the thing about everything in this article. None of these tells involve a single bad word. You can commit every one of them using nothing but plain, clean, list-approved vocabulary. The “not just X but Y” reflex, the rule of three, the hedges, the gerunds, the “which” chains, the weak endings, all of them are made of perfectly innocent words arranged in a machine way.

That is why the sentence layer is the real first line of defense and the word layer is just the screen door in front of it. A reader who has developed an ear is not listening for “delve.” They are listening for the rhythm, and the rhythm is built at the sentence level out of these constructions. You can hand that reader a piece with flawless vocabulary and they will still know, in about two paragraphs, because the music is wrong and the music lives here.

Fixing the sentence layer is genuinely hard, harder than the word layer by a wide margin, because there is no list and the fixes require an ear you have to develop. But it is also where the prose starts to actually sound human, because human rhythm is varied and surprising and these fixes all push toward variation and surprise. Get the sentences right and you have done real work that the careful reader will feel.

One more thing about this layer, because it is the layer where the skill of reading aloud earns its keep. Every tell in this article can be caught by ear, and most of them are nearly impossible to catch any other way. You will read past a “not just X but Y” silently a hundred times and never register it, because on the page it looks fine. Read it aloud and the construction announces itself, because your mouth feels the fake pivot in the middle. The same is true for the rule of three, for the hedges, for the weak endings. They are rhythmic problems, and rhythm is an auditory thing, so the eye skips them and the ear catches them.

This is the single most useful habit you can build for the sentence layer: read your work out loud, slowly, all of it. Not in your head, out loud, with your actual voice, because the voice catches what the eye forgives. Where you stumble, there is a problem. Where the rhythm goes flat and even, there is a cluster of these constructions. Where a sentence dies on a weak word, your voice will drop and trail off and you will hear it die. The ear is the instrument for this entire layer, and most writers never use it, because reading aloud feels slow and a little ridiculous. It is slow. It is the difference between catching the machine and not.

And you will still not be done, because clean sentences can assemble into machine-shaped paragraphs, which is the next layer down and the subject of part three.

Next in the series: Part three, the paragraph level, the outline wearing prose. Or go back to part one or the series hub.