AI writing tells: scene and voice

This is part four of a five-part series on AI writing tells, working down from the surface toward the spine. The first three parts handled the words, the sentences, and the paragraphs. Those layers apply to all prose, an essay, a blog post, a report. This layer is different. It only shows up when a writer stops explaining and starts rendering, when the job is no longer to convey a point but to put the reader inside an experience. Fiction lives here. So does memoir, narrative nonfiction, anything with a scene in it. And this is where the machine’s deepest limitation surfaces, because rendering experience requires having had some, and the machine never has.

The thing the machine cannot draw on

Everything in the first three parts could be framed as a habit, a tic, a pattern you could name and fix. This layer is not really a habit. It is an absence. The machine has never been cold, never been afraid, never stood over a body, never wanted someone it could not have, never felt a room go wrong. It has read millions of descriptions of all those things, and from those descriptions it has learned the words people use. What it has not got is the thing the words point at.

So when the machine writes a scene, it assembles it out of other descriptions of scenes. The result is prose that is about experience without being made of it, and the tells at this layer are all the specific ways that secondhand construction shows through. A reader who has lived feels the difference immediately, often without being able to say why. The why is that the machine is describing the map and has never walked the territory.

Let me name the specific tells, because vague talk about soul and experience helps nobody. The absence shows up in concrete, catchable ways.

The retreat into narration

Before the specific scene tells, the largest one, the habit that shapes everything else at this level. The machine does not want to write dialogue. Left to itself it retreats into narration, endlessly, and it will tell you what two people said rather than let them say it. Ask it for a tense confrontation and you get a paragraph reporting that one character told the other that they disagreed, and that the other replied that they understood but could not change their position, and the whole exchange arrives flattened into told-her-that and he-replied-that, with no quotation marks anywhere, because the machine is far more comfortable summarizing speech than staging it.

This is worth understanding, because it is one of the most reliable tells there is and one of the easiest to spot. Open a machine-written chapter and look at the proportion of dialogue to narration. A human writer of fiction leans hard on dialogue, because dialogue is where character lives, where conflict happens in real time, where the reader gets to watch instead of being told. The machine inverts this. Its default is a dense wall of narration with dialogue appearing rarely if at all, and when dialogue does appear it is brief and quickly swallowed back into the narration that the machine clearly prefers. The book reads like a long synopsis of a book, a report on events rather than a dramatization of them.

The reason is the same reason behind the clean-dialogue problem below. Real dialogue requires distinct voices, and the machine has one voice. Two characters talking means two different ways of speaking, two rhythms, two vocabularies, two people who want different things and pursue them through how they talk, and the machine cannot reliably produce that, so it avoids the situation that would expose it. Narration lets it stay in its single voice indefinitely. Every line of reported speech is a line it did not have to dramatize, a scene it did not have to stage, a moment where it chose the safe summary over the risky scene. The retreat into narration is the machine playing to the one thing it can do, continuous single-voiced prose, and avoiding the thing it cannot, two people alive on a page at once.

The fix is to convert. Go through the draft and find every place a scene was reported instead of dramatized, every told-her-that that should have been a line of dialogue, and stage it. Let the characters speak. Make their voices different. This is laborious, because it is essentially writing the scenes the machine declined to write, but it is the difference between a synopsis and a story, and there is no shortcut, because the machine avoided dialogue precisely because dialogue is the part it cannot do.

Characters who watch themselves

This is the signature tell of machine fiction, the one I would point to first. AI characters observe themselves from the outside, as if a camera floated a few feet away. “She found herself reaching for the phone.” “He realized he was holding his breath.” “She noticed her hands were shaking.” “He felt himself tense.” The character is not inside the experience. The character is watching a report of the experience arrive.

Real people are not spectators of themselves. When you reach for the phone, you reach for the phone, you do not “find yourself” reaching for it. When you are afraid, your hands shake and you are busy being afraid, you are not calmly noticing that your hands have begun to shake. The machine writes the noticing because it learned that interiority means reporting on the self, and so it stations the character outside the body, narrating the body’s actions like a sportscaster.

The reason the machine does this is the same distrust that runs through the whole series, pointed now at experience. It does not trust the action to convey the feeling, so it adds a layer of the character observing the action and labeling it. “His hands shook” is the human version, and it trusts you to know that shaking hands mean fear. “He noticed his hands were shaking, a sign of how afraid he had become” is the machine version, which does the action, then observes the action, then explains the observation, three layers where one would do, because it cannot let the bare physical fact stand and mean something on its own.

The fix is to collapse the character back into the body. Cut every instance of “found herself,” “realized he was,” “noticed that she,” “felt himself.” Write the action directly. “Her hands shook.” Then stop. Do not have the character observe it and do not explain what it means. The reader has hands. The reader has been afraid. They know. The moment you stop letting characters watch themselves, the prose drops from the floating-camera distance down into the actual body, and the scene goes from described to inhabited.

Dialogue where everyone is articulate

Machine dialogue is too clean. Every character speaks in complete, grammatical, well-formed sentences. Each finishes their thought before the next person starts. Each responds directly to what was just said. Questions get answered. Statements get acknowledged. The conversation proceeds in tidy, alternating turns, like a transcript of a debate between two unusually composed people.

Real conversation is a mess. People interrupt. They trail off and never finish the sentence. They answer a question with a different question, or with a non sequitur, or by ignoring it entirely and saying the thing they were already going to say. They talk past each other, each responding to what they assumed the other meant. They repeat themselves, fumble for words, start over, say something and then immediately walk it back. Two people having an emotionally charged conversation are barely communicating half the time, and the gap between what is said and what is meant is where all the real content lives.

The machine produces none of this raggedness, because it learned dialogue from written dialogue, which is already cleaner than speech, and because clean exchange is easier to generate than the broken, overlapping, evasive texture of real talk. So machine dialogue has a smoothness that reads as wrong to anyone paying attention, a sense that these are not people talking but positions being articulated.

The fix is to rough it up deliberately and specifically. Let a character cut another off mid-sentence. Let someone fail to answer the question and change the subject instead. Let a line trail into nothing. Most important, build the gap between the spoken and the meant. Two people having a hard conversation are usually having two conversations, the one out loud and the one underneath, and the underneath one is the real one. When your dialogue has that doubled quality, the said thing and the avoided thing running at the same time, it stops sounding like a transcript and starts sounding like people, who almost never say the thing directly.

The body’s small repertoire

Machine characters express emotion through a startlingly limited set of physical gestures, the same handful recycled across every scene. They nod. They swallow. They let out breaths they did not know they were holding. They feel their chests tighten. They quirk eyebrows. Something flickers across their faces. Their stomachs drop. The machine has a little kit of about a dozen body-language moves, and it deploys them over and over as the universal signals for nervous, sad, surprised, afraid.

The tell here is both the smallness of the repertoire and the vagueness of it. “Something shifted in his expression.” What thing? “She felt a chill run down her spine.” Every machine character feels that identical chill in that identical spine. The gestures are generic because they are not drawn from a specific body in a specific moment, they are drawn from the average of ten thousand described gestures, and the average of all bodies is no body in particular.

A human writer reaches for the specific, slightly wrong, unrepeatable physical detail. Not “she swallowed nervously” but the particular thing this particular person does, picks at the label on the bottle, talks too fast, goes very still, laughs at the wrong moment. The specificity is the realness. The fix at this layer is to ban the standard kit entirely, no held breaths, no quirked eyebrows, no chills down spines, no dropped stomachs, and force yourself to find the actual physical behavior of this actual character in this actual moment. When you cannot fall back on the generic gesture, you have to observe, even if you are observing something imagined, and the observed detail is always stranger and truer than the kit.

Everything is load-bearing

Here is a subtler scene-level tell, and a deep one. In machine prose, every detail matters. If an object is mentioned, it will return. If a trait is established, it will pay off. If the writer notes the weather, the weather means something. The machine includes only details that serve a function, because it learned that good writing has no wasted elements, and it applied that lesson too literally, producing a world where everything is significant and therefore nothing feels real.

Real life is full of details that go nowhere. A real writer, rendering a real-feeling scene, includes things that do not pay off, a passing observation, an object that is just there, a person mentioned once who never returns, a fact about the room that means nothing. This incidental texture is what makes a fictional world feel inhabited rather than constructed. The reader senses, without thinking about it, that the world extends past the edges of the plot, that it contains more than the story needs, because real worlds do.

The machine cannot do this naturally, because to the machine a detail with no function is an error. So its scenes have a curious airlessness, an engineered quality, where every element clicks into a slot and the reader feels, somehow, managed. The fix is counterintuitive: deliberately include details that do not matter. Not random noise, but the kind of incidental specificity a real observer would notice, the wrong-colored car, the crack in the ceiling, the name of a song playing that never comes up again. Let the world spill past the plot. The excess is what makes it feel real, and the machine’s refusal to waste anything is what gives it away.

The senses that never get past sight and sound

Machine scenes are filmed, not lived. They render what a camera and a microphone would catch, the look of things and the sound of things, and they almost never reach the senses a camera cannot capture. Smell, taste, temperature, the body’s sense of its own position and weight and fatigue, the senses that actually locate a person inside a moment, are largely absent from machine prose, because the machine learned scene-writing mostly from the visual and the audible, which dominate written description.

This is a quiet but powerful tell. A real scene has a temperature. The cold of an empty parking lot at two in the morning is different from the cold of a hospital corridor, and a writer who has been in both knows the difference and can put it on the page. A real scene has smell, which is the sense most wired to memory and the one that drops a reader into a place faster than any visual. The machine writes “the room was cold and quiet.” A human writes the specific cold, the radiator that clanks but does not warm, the smell of old coffee and floor cleaner, the way the chair is cold through your clothes when you finally sit. One is footage. The other is presence.

The fix is to push past sight and sound on purpose, in every scene. After you have drafted what the place looks and sounds like, ask what it smells like, what the air feels like on the skin, what the temperature is doing, what the character’s body is registering, the ache in the legs, the hollow of hunger, the particular tiredness that sits behind the eyes. You do not need all of them. You need more than the two the machine defaults to. The senses beyond sight and sound are the ones that prove a body was there, and the machine has no body, which is exactly why it forgets them.

The scene that runs at one speed

A real scene changes speed. It slows down and lives inside a single gesture when the moment matters, stretching three seconds across a paragraph, and it speeds up and covers an hour in a sentence when the ground just needs to pass. The variation in speed is how a writer tells the reader what matters, where to lean in and where to coast. The machine runs every scene at one speed, the same level of detail and the same granularity from the first line to the last, because it has no instinct for which moments deserve to be slowed into and which should be hurried through.

This produces a peculiar flatness even in scenes where the events are dramatic. A fight, a confession, a death, all rendered at the same even pace as a character making coffee, because the machine gives each beat its standard allotment of detail regardless of weight. The reader feels the flatness as a lack of emphasis, a sense that the scene does not know what is important in itself, which is exactly the case.

The fix requires you to decide, for each scene, where the moment is, the single beat the whole scene exists to deliver, and then slow violently into that beat while compressing everything around it. The walk to the door is a sentence. The hand on the doorknob, if that is the moment, is a paragraph. When the speed varies with the weight, the scene gains a shape, a building and a release, and the reader’s attention rises and falls with the writer’s, which is the opposite of the machine’s flat, uniform crawl.

The emotion explained after it is shown

This one bridges into the deepest layer, so it is a good place to end. The machine shows an emotion and then, not trusting the showing, explains it. “She slammed the door. She was furious.” “He stared at the floor and said nothing, overwhelmed by grief.” The scene does its job, the slammed door, the silence, and then a sentence arrives to label the feeling the scene already conveyed, in case you missed it.

You did not miss it. The slammed door was the fury. The label adds nothing except the machine’s anxiety that the rendering was not enough. This is the show-then-tell pattern, and it is one of the most reliable tells in all of machine prose, because the machine almost cannot stop itself from explaining, from putting the feeling into a clean abstract word after it has already put the feeling into action.

The human writer shows, and then trusts the showing, and stops. The slammed door stands alone. The silence stands alone. The reader supplies the fury and the grief, and because the reader supplied it, the reader feels it, which is the entire point. The instant you explain the emotion, you take it away from the reader and hand them a label instead, and a label cannot be felt. The fix is to find every place where a shown emotion is followed by a told one and delete the told one. It will feel risky every time. It is not risky. The showing was always enough, and the trusting is the thing that makes it land.

Where this leads

Notice what keeps surfacing as we go deeper. At the word level the machine could not trust the plain word. At the sentence level it could not trust the flat declarative. At the paragraph level it could not trust the reader to hold the point. And here at the scene level it cannot trust the action to carry the feeling, so it watches, labels, and explains at every turn. The same distrust, going deeper each layer, manifesting as a refusal to let anything stand on its own and mean something.

At this layer the distrust is aimed at the most important thing of all, the reader’s capacity to feel. The machine over-explains experience because it does not believe the reader will feel what it does not name. That single failure of trust produces the watching characters, the explained emotions, the load-bearing details, the labeled gestures, all of it, an entire layer of tells flowing from one refusal to let the reader do the reader’s job.

There is also a harder truth under this layer, and it is worth saying plainly because it shapes what you can and cannot fix. Some of the scene-level absence is not a habit the machine could drop. It is a consequence of the machine never having been alive. A writer fixes the watching characters and the labeled emotions and the missing senses by drawing on their own experience of being a body in a frightening room, and the machine has no such well to draw on. So when you take a machine scene and make it live, you are not removing the machine’s tells so much as pouring your own life into the hollow where a life should have been. The fix at this layer is not really editing. It is transfusion. You are giving the scene the lived experience it never had, because you have one and it does not.

This is why the scene layer is the first place where AI cleanup stops being subtraction and becomes addition. The earlier layers were mostly about removing, cutting the bad word, breaking the run-on, deleting the restatement. Here you have to add, to bring something to the page that was never in the draft because it could never be, the specific cold, the real gesture, the doubled conversation, the sense from inside a body. You cannot scrub a machine scene into a human one. You have to inhabit it, which means the writer has to show up and spend something, and that spending is exactly what the machine cannot do and you can.

And that refusal to spend has a bottom. There is a place where the distrust stops being a matter of words or sentences or even scenes and becomes the literal architecture of the whole work, the shape of the story itself built to never let anything cost what it should cost, never let anything stay broken, never make the reader sit with anything real. That is the spine, the deepest tell, the one you cannot reach with any edit because it is not in the prose at all. That is part five, and it is where the series has been heading the whole way down.

Next in the series: Part five, the spine, the deepest tell and the one you cannot edit away. Or return to part three or the series hub.