Why AI-Written Dialogue Sounds Off

Feed a scene to a machine and read the dialogue it hands back. Something is wrong, and you feel it a full beat before you can explain it. The lines are grammatical. They are on topic. Every character answers the question they were asked. And nobody in the room is a person. It reads like a very polite transcript of people who have never been nervous, never dodged a question, never said one thing while meaning its opposite. Let me name exactly what the machine keeps missing, because once you can hear it, you can never write it by accident again.

Real people do not answer the question

Here is the first tell. Machine dialogue is honest to a fault. Ask a character how they are, and the machine has them tell you how they are. But real conversation is mostly evasion. We answer the question we wish we had been asked. We change the subject when it gets close. We say "I'm fine" with a jaw so tight it means the opposite. A wife asks her husband if he called the doctor and he starts talking about the weather, and in that swerve the whole marriage tells on itself. The machine writes the man who calmly says yes or no. It has never had a thing it needed to avoid saying out loud, so it does not know that avoidance is where the scene lives. This is the same engine that drives show, don't tell, and the machine fails it in the mouth.

Subtext is a wound the machine does not have

Great dialogue runs on the gap between what is said and what is meant. "I'm with you now, Pop" is four flat words carrying a son's whole surrender. The machine can produce the four words. It cannot load them, because loading a line requires a lifetime of unspoken things pressing up under a small sentence, and the machine has no unspoken things. It has no back rooms. Everything it knows, it knows on the surface, evenly, at once. So its dialogue is all text and no undertow. It says exactly what it says, which is precisely why it sounds off, and precisely why the emotional truth of a scene keeps slipping through its fingers.

People rarely say what they mean. The machine cannot help saying exactly what it means, which is why no one believes it.Wr. Sarkhedi

The mess is the music

Listen to a real argument in an Ahmedabad kitchen and it is a disaster of overlaps, false starts, a point abandoned mid-sentence because a better wound opened up. People interrupt. They repeat themselves. They lose the thread and grab a worse one. That mess is not noise to be cleaned. It is the music of how humans actually collide. The machine, trained to be clear and helpful, sands all of it off. It gives you a tidy debate where two well-behaved positions take turns. Craft sites like No Film School and ScreenCraft keep teaching writers to break their dialogue on purpose, to let it stutter and swerve, and the reason is that clean dialogue is dead dialogue. The machine's instinct for tidiness is exactly the instinct that kills a scene.

Voices that are all the same person

There is a subtler tell too. In machine scenes, everyone talks alike. The gangster and the grandmother and the physicist all use the same balanced, slightly formal register, because the model is drawing on one averaged sense of how sentences go. But a real character has a fingerprint you could identify with the names stripped off. One person hides behind jokes. One over-explains when he is guilty. One uses ten words where five would do because silence frightens her. Distinct voices come from distinct inner lives, and the machine has one inner life, which is to say none. Everything it writes sounds like the same competent stranger doing every part, which is exactly the failure I unpack from the other direction in writing a character voice that does not sound like anyone else in the script.

A scene I wrote just to test this

Let me show you rather than tell you, because the difference is easiest to feel side by side. Here is a version of a scene a machine might produce, a father asking his son why he dropped out of college.

Father: Why did you leave your studies?
Son: I felt lost and unmotivated, and I did not think the course was right for me.
Father: I am disappointed, but I want you to be happy.An invented, illustrative exchange, written to fail on purpose

Every line is true, complete, and grammatically flawless, and the scene is dead the moment you read it, because both men are explaining themselves like witnesses in a deposition. Now here is the same beat, written the way I would actually build it, with the son unable to admit the real reason and the father unwilling to ask it directly.

Father: Your mother stitched three blouses extra this month for that admission fee.
Son: I know what it cost.
Father: Then you know what I am not asking you.Same scene, rewritten to carry what neither man will say

Nobody states a feeling in the second version. The father never says I am ashamed of you, and the son never says I was drowning and too scared to tell anyone. But the shame and the drowning are both fully present, sitting in the gap between the sentences. That gap is the entire difference between the two versions, and it is the gap a language model has no way to build, because building it requires having sat in a real kitchen where a real number, a real cost, stood in for a feeling nobody could say out loud.

Five fixes for a scene that reads machine-made

Whether the dialogue in front of you came from a model or from your own tired first draft, the repair is the same. Run these in order and the scene usually comes back to life.

1. Delete the direct answer. Wherever a character answers the actual question, make them answer a different one instead, and let the real answer leak through what they chose to talk about.
2. Give one character something to protect. A secret, a debt, a fear of being seen as weak. Every line they speak should now be quietly managing that exposure, even when the topic is dinner.
3. Break the rhythm. Let someone interrupt, lose the thread, or answer three exchanges late, the way real memory catches up mid conversation.
4. Strip the register to one person's habits. Read the scene aloud with the names covered. If you cannot tell who is speaking, rewrite until you can.
5. Find the object. A cup, a plate, a specific rupee amount. Let the object carry the freight the sentence is avoiding, the way the college fee does above.

A machine will always answer the question you asked it. A person answers the question underneath it, if they answer at all.

How to hear the fake note, and fix your own

Once you know the tells, you can audit any scene, human or machine, in a minute. Ask these:

• Is anyone avoiding something? If every line answers the question directly, the scene is probably dead.
• Is there a gap between said and meant? If the dialogue means exactly what it says, there is no subtext to feel.
• Could you tell who is speaking with the names removed? If not, you have one voice wearing four costumes.
• Is it too clean? Real speech stumbles. Perfectly balanced lines are the sound of no one.

Write the swerve, not the answer. The truth of a scene hides in what a person will not say.

Why this is the machine's permanent weakness

This is not a flaw the next version patches. It runs deeper than training data. Dialogue sounds true when it leaks true feeling sideways, through evasion and mess and the pressure of things a person cannot say, and all of that requires a person who has felt those things and learned to hide them. The machine has never needed to hide anything, because it has never felt anything worth hiding. That is why its dialogue will keep sounding off no matter how fluent it gets, and it is the same reason the harder crafts, the felt ones, stay human even as the mechanical ones fall to the tool.

I think about this in terms of debt. Every human being carries an emotional debt ledger nobody else can see, favours owed, insults never repaid, love confessed too late or not at all. When two people talk, the ledger is quietly being adjusted underneath every sentence, and the dialogue that moves us is the dialogue that lets us feel the ledger shift without ever seeing the numbers. A model has no ledger. It has no unpaid debts, because it has never owed anyone anything. Ask it to write a scene between two brothers who have not spoken in a decade, and it will produce a competent reconciliation. It will not know, in its bones, what it costs a man to say sorry to someone he has spent ten years being right about. That cost is the whole scene, and it can only be written by someone who has paid a version of it themselves.

This is also why the fix for machine-flat dialogue and the fix for a beginner's flat dialogue are the same fix, aimed at the same wound. Beginners write on the nose for the same reason the model does: it feels safer to say the thing plainly than to trust an audience to feel it in the gap. The machine has an excuse, it has never had a self to hide. You do not have that excuse, and that is good news, because it means the fix was always available to you and never to it. The new writers most exposed to this are the ones just starting out, which is a real and separate problem I wrote about in the piece on breaking in. For now, the practical lesson is small and permanent. Go listen to real people. Steal their swerves. The machine cannot, because it has never had a reason to look away, and it never will, because looking away requires having something to hide, and having something to hide requires having lived.

Common questions

Why does AI dialogue always sound the same no matter the character?

Because the model draws every line from one averaged sense of how sentences go. It has no distinct inner life for the gangster versus the grandmother, so both get the same balanced, slightly formal register. Distinct voices come from distinct wounds and habits, and the machine has none of its own.

Can you fix AI generated dialogue instead of rewriting it from scratch?

Sometimes, if you treat the output as a rough transcript rather than a finished line. Strip out the direct answers, add one evasion per exchange, and cut any sentence that states a feeling outright. You are doing the human work the draft skipped, not polishing what is already there.

Is on-the-nose dialogue the same problem as AI-sounding dialogue?

They overlap but are not identical. On-the-nose dialogue is a craft mistake any writer can make, and I go through the full repair for it in on-the-nose dialogue, how to spot it and kill it. AI dialogue sounds off for a deeper reason: it has never had anything to hide, so even when it avoids stating a feeling directly, the evasion itself feels invented rather than lived.

How do I train my ear to write dialogue that does not sound artificial?

Listen to real people argue, apologise, and flirt, and write down the exact sentence they used instead of the sentence you expected. The gap between what you expected and what they actually said is where a real voice lives, and it is the gap the machine cannot guess.