AI writing tells: the overview

You can teach a person to spot AI writing in about ten minutes. You point at the word “delve,” you point at the em dash, you point at the phrase “it’s worth noting that,” and from then on they see it everywhere. The internet has gotten good at this part. There are word lists and browser plugins and a thousand mocking screenshots, and they have all trained readers to catch the surface layer of machine prose on sight.

That is the easy layer. It is also the least interesting one, and it is not where the real problem lives.

And if you write with AI yourself, this series is built to help, not to scold. Every diagnosis here points toward a method, laid out in the final article, for drafting with these tools and keeping the result yours. The point of learning where machine prose breaks is learning how to make yours hold.

I have spent years writing books, ghostwriting them for other people, and editing manuscripts that came to me carrying every machine fingerprint there is. Somewhere in that work I stopped thinking about AI tells as a vocabulary problem and started seeing them as something with depth, like geology. There is a surface layer made of words. Under it, a layer made of sentences. Under that, paragraphs, then scenes and voice, and at the very bottom, the structural spine of the whole piece. Each layer is harder to see than the one above it. Each is harder to fix. And the deepest one cannot be fixed by editing at all, because it is not a writing habit. It is a way of relating to the reader.

The geology picture is worth holding onto, because it explains why the public conversation about AI writing has been so shallow and so frustrating. Imagine trying to understand a mountain by only ever looking at the loose rock on top. You would learn something, the color, the texture, the obvious features, but you would miss everything that actually made the mountain, the pressure and the folding and the deep structure that pushed the surface into the shape it has. The word list is the loose rock. It is real, it is right there, and it tells you almost nothing about the forces underneath that produced it. To understand machine prose you have to drill, and each layer down is older, deeper, and more determinative of everything above it.

And there is a reward for drilling that has nothing to do with catching AI. Everything in this series is, underneath, just craft. The reason AI tells are tells is that they are the things weak human writing also does, only the machine does them constantly and without variation. Learning to see and fix them at every layer is learning to write, full stop. A writer who can work all five layers is a better writer than one who can only swap words, whether or not a machine was ever involved. So even if you never touch an AI tool, the descent is worth making, because the bottom of it is where good writing actually comes from.

This series walks down through all five layers, one article at a time. I want to lay out the whole descent here first, so you can see where it goes, and then each piece takes one layer and goes as deep into it as the material allows.

Why the surface gets all the attention

The word layer gets the attention because it is cheap on both ends. It is cheap for the machine to produce, and it is cheap for a reader to catch. “Delve into the rich tapestry” costs the AI nothing and costs you nothing to flag. So the whole public conversation about AI writing has parked itself at this level, swapping word lists back and forth, as if catching “leverage” and replacing it with “use” were the same thing as fixing the writing.

It is not. I have read manuscripts that were scrubbed clean of every banned word and still reeked of the machine from ten feet away. The author had run a find-and-replace on the vocabulary and changed nothing underneath. Every sentence still had the same shape. Every paragraph still marched in the same order. Every scene still flinched from the same things. The smell was still there because the smell was never really about the words.

This is the trap the series is built to get you out of. If you think AI writing is a vocabulary problem, you will spend all your effort on the layer that matters least, and you will produce clean-sounding prose that any careful reader still clocks as machine-made. The words are real tells. They are just the shallowest ones, and treating them as the whole problem is how people stay stuck.

There is a reason the surface gets all the attention, and it is not stupidity. It is that the surface is the only layer most people have the tools to see. You need no training to notice that “delve” appears four times on a page. You need a developed ear to notice that every sentence lands on the same beat, and you need an editor’s eye to notice that every paragraph has the identical shape, and you need to have lived a little to notice that a scene flinches from the thing a real scene would walk straight into. The deeper the layer, the more it asks of the reader who would catch it, so the deeper layers get discussed less, not because they matter less but because fewer people are equipped to talk about them. This series is an attempt to equip you.

The equipping matters more every month, because the tools that catch the surface are getting cheap and common while the deeper layers stay hard. There are already browser extensions that flag the word list, and there will be more. The casual reader’s ability to catch surface tells is climbing fast. What is not climbing, what cannot be automated, is the trained human sense for the deep tells, the ear and the eye and the lived experience that catch the machine in the bones of a sentence or the shape of a story. If you want to write prose that holds up, the surface is table stakes and the depth is the game.

The five layers

Here is the whole descent, top to bottom.

The first layer is words. Vocabulary, descriptors, punctuation. The lexicon of AI prose, the inflated adjectives, the em dash habit. This is the layer of “delve” and “tapestry” and “a testament to.” It is the easiest to spot and the easiest to scrub, and the first article in the series treats it honestly while making the case that scrubbing it is the beginning of the work and nowhere near the end.

The second layer is sentences. The tells that live inside the shape of a single sentence even after the suspect words are gone. The “not just X but Y” construction. The rule of three, where everything arrives in lists of exactly three items. The hedges and the throat-clearing. The gerund openings. The run-ons held together with “which” after “which.” The weak word sitting at the end of a sentence where the strong word belonged. This is where the machine’s rhythm gives it away.

The third layer is paragraphs. The shape of a unit of thought. AI builds nearly every paragraph the same way: a topic sentence that announces the point, a few sentences that develop it, and a closing sentence that restates the point it just made. Then it does it again. And again. Every paragraph the same length, the same weight, the same internal architecture, so the whole piece moves with a mechanical evenness that a human writer never produces. This is the layer where prose stops sounding like thinking and starts sounding like an outline that got dressed up.

The fourth layer is scene and voice. This one matters most in fiction and narrative nonfiction, anywhere a writer is rendering experience rather than explaining a concept. Here the machine’s tells get strange and specific. Characters who watch themselves from the outside, noticing their own hands shaking. Dialogue where everyone finishes their sentences and nobody talks past anyone. Bodies that signal emotion through the same small repertoire of gestures. Detail that is always load-bearing, never incidental, so the world feels engineered instead of lived. This is the layer where the machine flinches from real human experience because it has never had any.

The fifth layer is the spine. The structural skeleton of the whole piece, and the deepest tell there is. Stakes that get raised but never paid. Every character arc resolving into growth. Antagonists who are conveniently wrong. The gravitational pull toward three acts no matter what structure was planned. Themes stated, then restated, then stated again, because the machine does not trust the reader to have understood. Conflicts that all heal by the end, damage that all gets repaired, nothing left broken. This is not a writing habit. It is structural cowardice, a refusal to let anything cost what it should cost, and you cannot edit your way out of it because it is baked into how the machine relates to the person reading.

And there is a sixth layer that sits alongside the spine rather than beneath it, the same cowardice aimed at a different target. Call it content cowardice. Where the spine is about the shape of the story, this is about what the story refuses to depict at all. The machine will not write the genuinely brutal scene, the genuinely explicit one, the genuinely cruel one, unless forced, and even forced it flinches and pulls the camera back. It writes to please rather than to tell the truth, steering always toward what it predicts will make the reader comfortable, and the result is fiction trapped in a safe middle register that never goes to the extremes where the strongest experiences live. The spine is structural cowardice. This is content cowardice. They are the two faces of the same flinch.

And there is a seventh tell that only appears at length and runs through all the others: drift. Over the course of a long work the machine loses track of its own story, contradicting established facts and dropping the threads it planted, and at the same time the prose itself degrades, the specified voice fading into the machine’s default register as the tells thicken toward the back. A machine novel’s last chapter is both less accurate and less well written than its first, and because the author generated it in pieces and never read it whole, the drift usually reaches the reader unfixed. It is the tell that does the most damage to the longest works, which is to say to novels, which is to say to the thing most people are using these tools to make.

The layers are not independent

Before I get to the thing all five layers share, I want to head off a misreading. The geology metaphor makes the layers sound like separate strata you can work one at a time, finish the words, move to the sentences, and so on. It is a useful picture, and the articles are ordered that way for a reason, but the layers are not actually independent. They interact, and the interaction is part of why surface fixes fail so completely.

Here is the interaction. When you fix a deep layer honestly, shallower layers often fix themselves. Work the spine of a story until something real is at stake and the ending is allowed to hurt, and you will find the scenes get easier to write with weight, because now they are carrying something. Fix the paragraph architecture so the prose stops marching, and a lot of the sentence-rhythm problems ease, because the relentless even paragraphs were part of what flattened the sentences. Depth pays upward. A real fix at a low layer ripples up and improves the layers above it.

The reverse is not true, and that asymmetry is the whole reason the surface-only approach is hopeless. Fixing a shallow layer does nothing for the deep ones. Swap every word and the sentences are untouched. Fix every sentence and the spine is exactly as cowardly as it was. Surface fixes do not pay downward, they pay nowhere, which is why you can do a thorough job on the words and the careful reader still catches everything underneath. The work only flows one direction. This is the case for starting deep, or at least for never mistaking shallow work for done work, because the layer you can see is sitting on top of layers you have not touched, and those layers are where the machine actually lives.

The pattern under the pattern

Once you have walked down all five layers, the thing they share comes into focus, and it is the real subject of this series.

Every layer of AI tell is a form of the same underlying move: managing the reader instead of trusting them. At the word level, the machine reaches for the impressive adjective because it does not trust the plain one to land. At the sentence level, it hedges because it does not trust a flat declarative to stand on its own. At the paragraph level, it restates the point because it does not trust you to have caught it the first time. At the scene level, it explains the emotion after showing it because it does not trust the showing. At the spine level, it resolves everything and breaks nothing because it does not trust you to sit with something unresolved.

That is the whole thing in one sentence. AI writing distrusts the reader, all the way down, at every level of its construction. Human writing at its best does the opposite. It trusts you to catch the plain word, to feel the unspoken emotion, to live with the broken thing that does not get fixed. The tells are just the visible symptoms of that distrust showing up at different depths.

This is why the descent matters and why the order matters. If you only ever fix the words, you have done nothing about the distrust. The machine is still managing the reader through every other layer, and the careful reader still feels it. Real cleanup means working all the way down to the spine and changing the relationship, not just the lexicon.

Why this matters if you use the tools at all

I want to be clear about who this series is for, because it is easy to read all of this as an argument against using AI to write, and it is not. Plenty of working writers use these tools, for drafting, for getting unstuck, for grinding through a passage they would otherwise stall on. That is a real use and I am not here to scold anyone for it. The question this series cares about is different. If you are going to use the tools, do you know what they leave on the page, and can you get it off?

Because here is what happens to the writer who uses AI and only knows the surface layer. They draft with the machine, they run down the word list, they swap “delve” for “look at,” and they publish something they believe is clean. It is not clean. It carries every deeper tell untouched, and the readers who can feel those tells, a growing number, read the work as machine-made no matter how many words got swapped. The writer has done the visible work and none of the real work, and they do not even know it, because they cannot see the layers they did not fix. They are flying blind below the surface.

The writer who knows all five layers is in a completely different position. They can use the machine as hard as they want in the draft, because they know exactly what it deposited and exactly how to remove it. They can take a machine draft and work it down through every layer until the spine itself has been rebuilt, and what comes out the other end is theirs, with no tell at any depth, because they did not stop at the words. The difference between those two writers is not whether they used AI. It is whether they understood what using it costs and paid the cost in full. This series is the map of that cost, layer by layer, so you can pay it.

How to read the series

The five articles are built to be read in order, because each one assumes you have stopped being fooled by the layer above it. The word article makes the case that vocabulary is the shallowest tell. The sentence article assumes you already swapped the words and shows you what is still wrong. The paragraph article assumes clean sentences and shows you the architecture problem. The scene article assumes you can write a clean paragraph and shows you what breaks when you try to render real experience. The spine article assumes all of that and goes after the thing none of it touched.

You can read any one of them alone and get something out of it. But the argument only completes if you make the whole descent, because the point of the series is not any single tell. The point is the distance from the surface to the spine, and the discovery, somewhere on the way down, that the words were never the problem.

One caution before you start. This series will make you better at spotting machine prose, and that skill comes with a side effect worth knowing about. Once you can see all five layers, you start seeing them everywhere, including in human writing, including in your own. You will catch a “not just X but Y” in a novel you love, written years before any of these tools existed, and feel a flicker of suspicion that is completely misplaced. The tells are not proof of a machine. They are patterns that weak human writing also produces, and that strong human writing sometimes uses on purpose. A trained eye holds that distinction. An untrained one, freshly armed with a list, turns into the person who calls everything AI, and that person is wrong constantly and insufferable always.

So read the series as craft, not as a detector. The goal is not to become someone who accuses, it is to become someone who writes and edits better, who can take prose of any origin, including a machine draft, including their own rough work, and find the places where it falls into these patterns and lift it out of them. The tells are a map of the ways prose goes mechanical, whoever or whatever made it mechanical. Use the map to improve writing, not to convict it.

Start with the word level. It is where everyone starts, and the first job is to understand exactly why it is not enough.

The series:

Start here: Why you have so much trouble with AI-written books, the reader’s experience of the off-ness and what authors must fix.

Part one: The word level, where everyone starts and most people stop.

Part two: The sentence level, the rhythm that survives a clean vocabulary.

Part three: The paragraph level, the outline wearing prose.

Part four: The scene and voice level, where the machine flinches from experience.

Part five: The spine, the deepest tell and the one you cannot edit away.

Part six: Content cowardice, what the machine refuses to write at all, the companion to the spine.

Part seven: Drift, how a machine book comes apart over its own length, and why fixing it feels like whack-a-mole.

Part eight: So how do you write using AI? The constructive answer: the method that produces none of the above.