AI writing tells: the word level

This is part one of a five-part series that walks down through the layers of AI writing tells, from the surface to the structural spine. The hub article lays out the whole descent. This piece takes the top layer, the words, and makes a case you might not expect from a piece about AI vocabulary: the words are real tells, and fixing them is almost worthless.

Let me explain that, because it sounds like a contradiction.

The layer everyone already knows

If you have spent any time online in the last couple of years, you already know the word list. “Delve” is the famous one. “Tapestry,” usually rich, usually woven. “A testament to.” “Navigate the complexities.” “Leverage,” “utilize,” “robust,” “seamless,” “holistic.” The inflated adjectives: “remarkable,” “profound,” “fascinating,” “striking.” The verbs that want to sound like a TED talk: “unlock,” “unleash,” “elevate,” “empower.” And the punctuation tell that launched a thousand arguments, the em dash, deployed three times a paragraph to create a false sense of rhythm.

People can spot these now. That is genuinely new. Two years ago the average reader could not reliably tell machine prose from human prose, and now a large number of them can flag it on sight, and the words are how they do it. So I am not going to pretend the vocabulary tells are fake. They are real. They are the most reliable surface signal there is. If a piece is dense with “delve” and “tapestry” and “testament to,” a machine was involved, and a reader who knows the list will know it.

Here is the problem. Knowing the list has convinced an entire generation of writers that AI cleanup means running down the list and swapping each word for a plain one. Delve becomes look at. Leverage becomes use. Utilize becomes use. Robust becomes strong. Em dashes become commas. And the writer believes that when the list is clean, the job is done.

The job is not done. The job has barely started. I have edited manuscripts that were spotless at the word level and unmistakably machine-made at every level below it, and the cleanness of the vocabulary actually made them worse, because it made the author think the prose was fixed when nothing underneath had moved.

Why the words are there in the first place

To understand why scrubbing words does so little, you have to understand why the machine reaches for those words to begin with. It is not random. Every one of those tells is the same impulse pointed at the vocabulary level.

The machine reaches for “delve” instead of “look at” because “delve” sounds more substantial. It reaches for “a testament to” instead of “shows” because it sounds more weighty and considered. It reaches for “remarkable” and “profound” because plain nouns feel naked to it and it wants to dress them. Every inflated word is the machine not trusting the simple version to carry the meaning, so it inflates.

That impulse, the distrust of the plain version, is the thing. It does not live in the words. It lives under them, and it shows up at every other layer too. When you swap “delve” for “look at,” you have removed one symptom of the impulse. You have not touched the impulse. The next sentence will hedge, the next paragraph will restate itself, the next scene will over-explain its emotion, all driven by the same distrust that produced “delve.” You cleaned one pixel. The image is unchanged.

This is why find-and-replace fails. It treats the words as the disease when they are a rash. You can suppress a rash and leave the infection completely intact, and that is exactly what a vocabulary scrub does.

The words you cannot find on any list

There is a second reason the list approach fails, and it is more practical. The famous words are only the famous words. The actual vocabulary layer of AI prose is much wider than any list, and most of it is made of words that are perfectly good words used in a particular machine way.

Take “journey.” Used to mean an actual trip, it is fine. Used to mean “the process of learning to bake bread” or “her journey toward self-acceptance,” it is an AI tell, and no plain synonym fixes it, because the problem is not the word, it is the reaching for an abstraction where a concrete thing belonged. Take “navigate.” A ship navigates. A person does not “navigate a difficult conversation” unless a machine wrote the sentence. Take “foster,” “cultivate,” “nurture,” all fine words, all deployed by the machine as soft abstractions that mean roughly “cause” but sound more tender and considered.

These do not show up on the mocking screenshots because individually they are innocent. You cannot ban “journey.” But in aggregate, this layer of soft abstraction is a louder tell than “delve” ever was, and you will never catch it with a word list, because there is no list. You catch it by understanding the move: the machine prefers the abstract, gentle, slightly inflated version of any word over the concrete, plain, specific one. Once you see the move, you see it everywhere, in words no list will ever contain.

Let me show you the move in a single sentence so it stops being abstract. The machine writes: “Building a brand is a journey that requires you to navigate challenges and foster meaningful connections with your audience.” Not a single word on the famous list. “Journey,” “navigate,” “foster,” “meaningful,” “connections,” all innocent, all clean if you are checking against a screenshot. And the sentence is pure machine, because every noun and verb in it is the soft, abstract, slightly inflated version of something concrete. What does the sentence actually say? “Building a brand takes time, you will hit problems, and you have to win people over.” That is the human version, and notice it is shorter, plainer, and says more. The machine version was longer and emptier, padded with soft abstractions that each gestured at a real thing without naming it. The tell was not any one word. It was the consistent preference for the gesture over the thing.

This is why the soft-abstraction layer is the real word-level tell and the famous list is just its loudest symptom. “Delve” is a soft abstraction that happens to be famous. “Navigate,” “foster,” “journey,” and a hundred others are soft abstractions that are not famous, and the machine reaches for all of them by the same impulse. If you only hunt the famous ones, you cut the loud symptom and leave the disease, and the disease is the preference itself, which keeps producing new soft abstractions faster than any list can name them.

The adverb problem

There is a whole category of word-level tell that lives in the adverbs, and it deserves its own treatment because it is so consistent. The machine over-modifies. It cannot let a verb or an adjective stand alone, so it props each one up with an adverb, and the adverbs are almost always the same small set of intensifiers and softeners.

On the intensifier side: “incredibly,” “remarkably,” “deeply,” “truly,” “genuinely,” “absolutely,” “completely.” The machine writes “incredibly important” where a human writes “important,” “deeply meaningful” where a human writes “meaningful” or, better, names the specific meaning. The intensifier is the machine trying to add weight it has not earned, and it adds it the lazy way, by bolting on an adverb instead of choosing a stronger word or building a stronger sentence. “Incredibly important” is weaker than “important,” not stronger, because the reader registers the reach for emphasis and discounts it. The intensifier announces that the writer did not trust the plain word to carry, which is the same distrust that drives every other word-level tell.

On the softener side: “somewhat,” “rather,” “quite,” “fairly,” “relatively,” “arguably,” “perhaps.” These hedge the claim, lowering its confidence, and the machine sprinkles them constantly because a flat assertion feels too bold to it. “This is arguably the best approach” is a machine sentence. “This is the best approach” is a human one, and if you are not sure it is the best approach, the fix is not “arguably,” the fix is to figure out what you actually think and say that. The softener is a way of making a claim without committing to it, and committed claims are one of the things that make human prose feel human.

The fix for the whole adverb layer is close to mechanical, and it is one of the most satisfying passes you can run. Go through and look at every adverb. For each one, ask whether the sentence is stronger without it. The honest answer is almost always yes. “Incredibly important” becomes “important.” “Somewhat difficult” becomes “difficult,” or, if it really was only a little difficult, becomes a specific description of how. “Truly remarkable” becomes whatever the specific remarkable thing actually was. When you strip the propping adverbs, two things happen: the surviving words have to carry their own weight, which exposes the weak ones so you can replace them, and the prose gains the confidence of someone who says what they mean without hedging or hyping. The adverbs were a crutch, and the writing walks better without them.

The em dash, and what the argument got wrong

The em dash deserves its own section, because the public argument about it has been mostly stupid and it is worth saying why.

The claim went around that em dashes are an AI tell, and the counter-claim went around that plenty of human writers love the em dash, which is true, and somewhere in the noise the actual point got lost. The point was never that em dashes are inhuman. Good writers have used them for centuries. The point is the frequency and the function.

A human writer who loves em dashes uses them for a specific effect, an interruption, a sharp aside, a sudden swerve, and uses them sparingly enough that the effect still lands. The machine uses them as connective tissue, three or four to a paragraph, to fake a rhythm it cannot produce with sentence structure. The human em dash is a choice. The machine em dash is a tic. The tell is not the punctuation mark. The tell is the density and the sameness of the function, the same little dramatic pause deployed over and over with no variation in why.

This is the whole series in miniature. The surface argument fixated on the mark itself, banned or not banned, human or not human. The real tell was one level down, in the rhythm the marks were faking. You can write a piece with zero em dashes that has the exact same broken rhythm, because the rhythm problem was never about the dashes. It was about the machine not knowing how to build a varied cadence out of the sentences themselves, which is the subject of part two.

What scrubbing words actually buys you

I want to be fair to the word layer, because I have spent most of this piece telling you it does not matter much, and that overstates it.

Scrubbing the vocabulary buys you one real thing: it gets you past the readers who only know the surface. A large number of people can catch “delve” and nothing deeper. For them, a clean vocabulary genuinely reads as human, and depending on what you are writing and who you are writing for, that might be enough to matter. A blog post for a general audience that has had its obvious tells removed will pass with most of its readers. That is a real benefit and I am not going to pretend it is nothing.

But it is a floor, not a ceiling. It gets you past the casual reader and stops dead at the careful one. Any reader who has developed an ear for the deeper layers, and there are more of them every month, will still clock the prose, because everything under the words is untouched. So the honest way to think about the word layer is this: scrub it, yes, because there is no reason to hand the casual reader an easy catch. But understand that you have cleared the lowest bar there is, and the writers who think clearing it means they are done are the ones producing the clean-sounding, soulless prose that the careful reader spots in a paragraph.

Why the word scrub feels so productive

There is a psychological trap in the word layer worth naming, because it is the reason so many writers stop here believing they are done. Scrubbing the vocabulary feels enormously productive. You can see the work happening. Every “delve” you catch and replace is a visible win, a concrete thing fixed, and you can watch the count of bad words drop toward zero as you go. The pass has a satisfying shape, a clear start and a clear finish, and at the end you have a document measurably different from the one you started with.

That feeling is the trap. The visible, countable, finishable quality of the word scrub is exactly what makes it feel like the whole job, when it is the smallest part of the job. The deeper layers do not give you that satisfaction. You cannot count the sentence-rhythm fixes the way you count the words, and you can never be sure you have caught them all, and the work has no clean finish line, just a gradual improvement you have to feel rather than measure. So the word layer feels like accomplishment and the deep layers feel like uncertainty, and human nature steers people toward the accomplishment and away from the uncertainty.

I have watched writers spend an hour gleefully hunting “delve” and “leverage,” finish with a clean word list, and declare the piece de-machined, when the prose underneath had not moved an inch. They did the fun, visible, finishable part and skipped the hard, invisible, unfinishable part, and they felt great about it, because the part they did produced a satisfying before-and-after and the part they skipped would have produced only a vague sense of better. If you take one thing from this article, let it be a suspicion of how good the word scrub feels. That good feeling is not the feeling of finishing. It is the feeling of doing the easy thing first and mistaking it for the whole thing.

How to actually fix the word layer

Since you should clean it anyway, here is how to do it in a way that does more than swap synonyms.

Do not run a find-and-replace. The find-and-replace produces the worst result, a piece where every “delve” became “look at” mechanically, which creates its own flat sameness. Instead, when you hit an inflated word, do not reach for its plain synonym. Reach for the specific thing you actually meant. “Delve into the causes” does not become “look at the causes.” It becomes “the three things that caused it,” because the specificity is the cure, not the smaller word. The machine reached for the abstract verb because it did not have a specific thing to say. Your job is to have the specific thing.

This is why the word layer, done right, forces you down into the layers below it. You cannot fix “navigate the complexities of grief” by swapping “navigate.” You have to know what specific complexity, and the moment you ask that question you have left the word layer and entered the real writing. The vocabulary tell was a flag planted on top of a hole where a specific thought should have been. Filling the hole is the work. Swapping the word just repaints the flag.

The cluster, not the word

A last point about the word layer, and the one that will make you better at reading it than any list could. The tell is rarely a single word. It is the cluster, the density of these reaches packed close together, and density is something a list cannot measure.

One “leverage” in a thousand words means nothing. A human might write it. But “leverage” and “robust” and “seamless” and “holistic” all within three sentences is a machine signature, not because any one of them is damning but because no human reaching for plain language would stack four inflated abstractions that tightly. The machine stacks them because the inflated register is its default, so once it is in that register it stays there, and the words clump. A human drops into plain language between the occasional fancy word, so even a human who writes “leverage” surrounds it with concrete, specific, grounded language, and the fancy word stands alone instead of in a cluster.

This is why reading for clusters beats reading for words. When you scan a piece, do not just flag the suspect terms, notice where they bunch. A paragraph with one soft abstraction is probably fine. A paragraph where every key noun and verb is a soft abstraction is machine prose regardless of whether any single word made your list. Train yourself to feel the density, the proportion of inflated-to-concrete language in a given stretch, and you will catch machine vocabulary that no individual word would have flagged, because you will be reading the register instead of the lexicon. The register is the real tell. The famous words are just where the register gets loud enough for anyone to hear it.

That is the whole case for part one. The words are the most visible tell and the least important one. Clean them, because there is no reason not to, but do not believe for a second that cleaning them fixed the writing. Everything that actually makes the prose feel machine-made is still sitting there underneath, in the sentences, which is where part two goes.

Next in the series: Part two, the sentence level, the rhythm that survives a clean vocabulary. Or return to the series hub.