Why Does AI Writing Give Its Characters Lines Instead of Voices?
This is Essay 20 of 22 in the series Why Choose 100% Human Authors, addressed to readers who want to understand what a human author does that a machine cannot and why that difference matters to the experience of literary fiction.
The Test
There is a test any writer can apply to dialogue. Read it aloud, and ask: would two human beings, in any century, ever say this to each other?
The question sounds simple. What it uncovers is not.
A language model generates dialogue by predicting what two people in a given situation would likely say. The result is exchanges that convey the information the scene appears to require at that moment. The characters hit their marks and do not, for a single sentence, sound like people.
This is the difference between dialogue that reports and dialogue that reveals. Reported dialogue exists to move information from the page into the reader’s head. It is the writer’s hand passing notes through the mouths of characters. Revealing dialogue does something harder: it shows who these people are, what they want, and what they will not say. The information may or may not arrive. What matters is what the dialogue reveals about the people speaking.
What Reported Dialogue Sounds Like
Consider a scene in which two brothers meet after their father’s death to discuss the family estate. A language model will produce the relevant information. Ownership, debt, sentiment, decision. It will distribute these items between the brothers so that neither carries all the exposition. The result is something like this:
“Father left the estate to both of us equally,” Matteo said. “But the debts are significant. We will need to decide whether to sell or to take on the burden together.”
“I had no idea the situation was so serious,” Luca replied. “You always handled these matters. What do you think we should do?”
The scene is not incoherent. The information arrives. But no two brothers have ever spoken this way after their father’s death—not in grief, not in accusation. What the language model has produced is a transfer of information wearing the costume of an exchange between siblings.
What the Human Writer Understands
Approaching the same scene, they know from experience or previous reading that the brothers are not only exchanging information. They are negotiating a relationship under pressure they cannot name, in the wake of a loss neither has finished carrying.
The written dialogue may contain none of the facts, or it may contain them sideways, in the wrong order, or interrupted:
“He left it to both of us,” Matteo said, and then did not say anything else for a long moment.
Luca looked at the window. “Both of us.”
“There are debts. I thought you knew.”
“How would I have known?”
This is a different scene. Not because it withholds information, but because the information is not what the scene is about. What the scene is about is legible in the repetition, the question that does not ask what it appears to ask. A reader understands that these brothers have a history the narrative has not explained. “I thought you knew” carries blame without stating it; “How would I have known?” is not a request for clarification.
None of this requires annotation. The writer trusts the reader to find what lies underneath the surface.
The Problem of Giving Characters Permission to Speak
Reported dialogue has another signature: the prompt-and-respond structure, in which one character asks a question whose only function is to license the other character’s explanation.
“What happened at the guild meeting?” she asked.
“The master refused to sponsor any of us,” he said. “He has his reasons, I suppose, but none of us expected it.”
This exchange is not wrong in the way that bad grammar is wrong. But it is constructed rather than spoken. The question does not come from curiosity or from character; it comes from the writer’s need to get the guild meeting into the scene. It is a mechanical hinge, and the reader feels it as such.
Human writers fall into this trap too. The difference is that they can hear it. When reading the draft aloud, they catch the question that sounds like a prompt and the answer that sounds like a briefing. They know what it means when a character says what the reader needs to hear at the moment the reader needs to hear it. They revise because they know that characters who exist to inform are hollow.
A language model does not hear this. It generates the exchange that satisfies the informational requirement and moves on. There is no subsequent draft in which the prompt-and-respond is replaced by something that sounds like people.
What the Language Model Cannot Do
The chief problem with reported dialogue is not that it sounds unnatural but that it has nothing at stake.
When two people speak to each other, something is always in the balance. Approval, position, love, the past, the story they each tell about themselves. Dialogue is not two people taking turns. It is two people pursuing separate aims through the same conversation.
A language model cannot create this because it cannot render interiority. It can generate a character who claims to be angry without the anger touching a single word spoken. It can generate a scene in which two people disagree and both articulate their positions with the even-handed precision of a debate exercise.
What it cannot generate is the conversation in which a woman says something she does not mean because she is afraid to say what she does mean, and the man she is speaking to hears what she said and not what she meant, and neither of them leaves the room knowing what the other one was saying. That scene requires a writer who has been in that room or near it, who knows what it costs to speak or to stay silent, and who can hold both costs in the same sentence without explaining either.
Rio de Janeiro, the iii day of July, MMXXVI.
