Between 70% and 90% of all human communication is nonverbal. Tone, posture, facial micro-expressions, the half-second pause before an answer -- these carry more meaning than the words themselves. Strip all of that away, and what remains is dialogue on a page: naked words, no body, no voice, no face.
That's the impossible job of fiction dialogue. It must do what an entire human body does -- but with typography alone.
Consider six lines:
"We need to talk."
"Okay."
"About yesterday."
"I know."
"What do you think?"
"I don't know."
Cover the character names. Try to identify which lines belong to the impulsive, emotional protagonist and which to his calm, repressed wife.
Impossible. Their voices have vanished. What's left is information ping-ponging between two empty shells -- A speaks, B responds, repeat. If the names disappeared and nobody could tell the speakers apart, the dialogue has failed at its most basic function.
Elmore Leonard put it best: "If it reads like writing, I rewrite it." He meant that artificial polish -- every sentence grammatically pristine, every exchange perfectly structured -- kills the illusion of real speech. The goal isn't recording what people say. The goal is manufacturing a very specific lie: words that feel spontaneous but carry ten times the weight of actual conversation.
Dialogue Is Not What People Say
Grab a recorder. Tape thirty seconds of lunch conversation with a friend. Transcribe it word for word.
"So... yeah, and then he was like... wait, no, that's not... okay so basically what happened was... um..."
Unreadable. Real speech is directionless. It repeats, wanders, breaks mid-thought, answers questions that weren't asked, ignores questions that were. It has no purpose beyond filling silence.
Fiction dialogue is a performance wearing the mask of spontaneity. Every word is selected. Every pause is engineered. And it does three things real conversation almost never does simultaneously.
It reveals character. Not through what is said but through how. A retired professor says "this is logically untenable." His seventeen-year-old grandson says "that's total BS." Same meaning. Completely different people. Vocabulary, sentence length, rhythm, attitude -- stack these differences and what emerges is voice.
It advances plot. A conversation that ends with no shift in relationship, information, emotion, or power balance has no reason to exist. Cut it. The story loses nothing.
It manufactures tension. Good dialogue is never harmonious information exchange. Two people sit at the same table carrying different agendas, different fears, different secrets. The tightness readers feel comes from the current running beneath the words -- what's happening under the surface.
This is why dialogue is brutally difficult. The writer isn't transcribing speech. The writer is constructing an illusion: something that sounds casual while every syllable is doing work.
The Nature of Voice
"Voice" is one of the most overused and least understood words in craft discussions.
Surface voice is easy to spot. One character speaks in clipped sentences. Another drops academic vocabulary into casual settings. A third ends every other statement with "right?" Word choice, sentence length, verbal tics -- these are the skin of voice.
The skeleton underneath is what matters.
Voice reflects how a person metabolizes the world.
A control-oriented person speaks with surgical precision. Every word vetted, nothing volunteered, nothing wasted. An anxious person speaks like someone hopping across stepping stones -- leaping from topic to topic, sentences cracking in half. A conflict-avoider drowns in hedging language: "maybe," "I guess," "we'll see," always leaving an exit. A dominator interrupts, hijacks, and fires rapid questions to corner the other person.
So the question when writing dialogue isn't "what words would this character use." The question sits deeper: facing this specific situation, what are they afraid of? What do they want out of this conversation? What are they hiding? What do they actually feel about the person across from them?
Those answers shape every word that leaves their mouth.
Slapping a verbal quirk onto a character isn't enough. Catchphrases are decoration. Dialects are costumes. Strip those away -- if two characters still sound identical, their voices are fake. Real voice grows from the psychological bedrock: fear, desire, history, scars.
In Slima's Writing Studio, each character can have a dedicated file for this psychological groundwork. Not for the reader to see -- for the writer to consult mid-scene, to remember who this person is before writing what they say.
What AI Dialogue Failures Teach
AI writes bad dialogue. And those failures are more instructive than any success could be.
Ask an AI Assistant to write an argument between two characters. Read the output. Nine times out of ten, both characters sound like the same person in different costumes. Both polite. Both answering questions fully. Neither interrupting. Neither saying something irrational in the heat of anger. Like two trained customer service representatives role-playing a conflict scenario.
The reason is structural. AI has no personality. It can simulate external markers of personality -- tone, diction, sentence patterns -- but it cannot inhabit a specific human being. Its dialogue gravitates toward the statistical average: the most common conversational patterns in its training data, not what one particular soul would say.
But this is precisely where it becomes a tool.
Take a character in progress. Feed the AI their background, fears, psychological profile. Ask it to speak as that character. Then listen -- not for what it gets right, but for what it gets wrong.
"No -- he wouldn't use that word."
"No -- under this kind of pressure, he doesn't open up. He shuts down."
"No -- he would never apologize first. His pride won't allow it."
Every "no" is a boundary discovered. It defines where this character's voice ends -- cross this line and it's no longer them. Write those boundaries down. That is the character voice rulebook. Not given by AI. Found by the writer in the act of rejecting AI's attempts.
Dialogue as Discovery
AI is not here to write dialogue for anyone. That needs to be clear.
It's here to help discover characters.
Open the AI Assistant in Slima's Writing Studio. Run an experiment: have the AI play a character from the manuscript, then talk to it. Not to produce copy-pasteable lines. To test, through improvised exchange, how deeply the character is actually understood.
Say the character is Marcus. Former soldier. Trusts nobody. Speaks like he's rationing words. Feed that to the AI. Start talking.
Sometimes the AI nails it. A response comes back terse and uncomfortable, and something clicks: "Yes. That's Marcus. That exact clipped discomfort." The confirmation sharpens the picture of who Marcus is.
Sometimes the AI misses. It makes Marcus eloquent, warm, metaphorical. Immediately wrong. Marcus doesn't do that. The rejection is equally valuable -- it maps territory Marcus would never enter.
The third case is the most revealing. The AI says something, and there's hesitation. "Would Marcus say this? I genuinely don't know." That uncertainty exposes an unexamined zone in the character -- a blurry patch the writer never confronted.
This is the real payoff of AI dialogue practice. Not usable lines. Not polished scenes. Something harder to get and more important: collision with the limits of understanding. Every "no" is a boundary confirmed. Every "I don't know" is a blank that needs filling.
Save those discoveries. Open a Snapshot in Slima to preserve each round's findings. When it's time to write the actual draft, those records become navigation.
Between What Is Said and What Isn't
Raymond Carver's characters almost never say what they feel. They discuss the weather. They argue about what to have for dinner. They ask how the day went. The real conversation happens underwater.
Picture a couple at the edge of breaking up. Dinner table. They're discussing whether to see a movie tomorrow. On the surface -- scheduling. Underneath every sentence: is this relationship still worth the effort? Who breaks first and says the word both are thinking but neither will speak?
This is subtext. The field where dialogue actually takes place.
AI writes subtext poorly. Its default is directness -- characters stating thoughts and feelings because that's the "most efficient" communication. "I think we should break up." "I'm sad." "I love you but we're not right for each other." Clear. Explicit. And completely dead on the page.
People don't talk like that. Especially not in the moments that matter most. In those moments, the words get stuck.
They hint. They dodge. They answer a question that wasn't asked. They say one thing and mean another. They ask something without wanting an answer. Or they say nothing at all -- and the silence itself hits harder than any sentence could.
When writing dialogue, three questions matter. What does this character actually want to say? Why won't they say it? Between what they choose to say and what they want to say -- how wide is the gap?
That gap is where tension lives. Readers feel it. They lean forward without knowing why, trying to decode what the words really mean beneath their surface. That gap turns dialogue from text into something alive, something that breathes.
Hearing Your Characters
Many writers describe the same phenomenon: past a certain point, a character's voice starts arriving on its own. No need to engineer what they'd say. Just go quiet and listen.
This isn't mysticism.
It's the result of sustained immersion. Months spent with a character -- knowing their most humiliating childhood memory, knowing what wakes them at 3 a.m., knowing the desire they'd never admit out loud. When that depth is reached, the voice grows from the details organically. No templates needed. No frameworks. The character simply speaks, and the writer takes dictation.
AI can accelerate this process, but not the way most people expect.
Don't use it to generate dialogue. Use it as a mirror.
Open the AI Assistant. Run a character interview. Have the AI ask questions, answer as the character. "When you were fifteen, who disappointed you the most?" "If you knew the world ended tomorrow, who would you visit tonight?" These questions force entry into the character's core. Not answering "what kind of person is he" in the abstract -- speaking in his voice, from inside his skin.
Or reverse it. Have the AI play the character. Ask questions. Watch which answers earn a nod, which earn a frown. The frowns matter more. Why does this feel wrong? What should the right answer sound like?
Each round is practice in closing distance with the character. After each session, open a Snapshot in Slima and save the voice rules discovered. Next time the draft reopens, those records serve as a compass.
Do this enough, and the voice starts arriving unbidden. Not magic. The writer finally knows this person well enough to hear them.
Dialogue isn't lines a writer assigns to characters. It's what characters say through the writer's hands. When that state is reached -- the character is alive.
Spend an hour in improvisational conversation with the AI. Play one character, let it play another. Most of the output won't be usable. That's not the point. The point is that somewhere in the exchange, layers of the character that were never written will surface.
Then reopen the manuscript. Delete every "I know" and "okay" -- those placeholder responses that could belong to anyone. Replace them with silence. With a subject change that reeks of avoidance. With a question that sounds unrelated but carries a blade of aggression underneath.
Not a single line names an emotion directly. But every line bleeds with it.
Do that, and even with the names covered, readers know exactly who's speaking.