Why Does AI Writing Produce Sentences That Have Never Been Heard?
This is Essay 14 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 That AI Cannot Pass
There is a test available to any writer, requiring no special equipment and no second opinion. Read the prose aloud. The sentence that cannot be spoken without stumbling has revealed itself. The rhythm that rises without resolving, the clause that arrives too late to carry its weight, the phrase that seemed inevitable on the screen and collapses in the mouth: all of these are exposed the moment the writer’s voice must carry them.
A language model has no voice. It has processed language as text—as sequence, as statistical relationships between one word and the next—but it has never heard a sentence land. It does not know that two long periodic sentences placed together without relief will make the reader’s eye slow before their mind has registered why. It produces prose by prediction, and prediction has no ear.
What Prose Without an Ear Sounds Like
The failures of deaf prose are not always dramatic. They accumulate. Here is a sentence a language model might produce:
The merchant, having considered the matter at length and having consulted with his associate, who had experience in such affairs, determined that the most prudent course of action available to him under the circumstances would be to delay.
The sentence is grammatical. Every clause connects to the next. But read it aloud and feel what it does: the string of participial phrases at the opening, the embedded relative clause, the long approach to the verb, the verb arriving at the end with no weight to carry because the sentence has spent all its energy on its own scaffolding. The reader reaches delay and feels not resolution but exhaustion.
A writer with an ear might make it longer. But in the act of writing or revising, they would feel that the sentence has lost the thread of itself, and they would rebuild it so that the weight falls where it should:
He consulted his associate. The man had experience in such affairs, and his counsel was not to act. The merchant waited.
The stress falls on waited—the last word, the thing the whole passage has been moving toward. The reader arrives there without effort, carried by the prose.
Rhythm That Neither Rises Nor Falls
The deeper problem in AI prose is not the individual sentence but the cumulative effect of sentences that have been produced without any sense of how they relate to one another in sound. Rhythm in prose is not meter; it is not the regular recurrence of stressed and unstressed syllables as in verse. It is the management of pace: the long sentence that slows the reader before the short one releases them, the period that gives weight to what follows.
A language model produces sentences at a consistent length and syntactical complexity because the patterns in its training data reward sentence-level competence. The result is prose that moves at a single speed. Consider a brief passage:
The city was filled with the sounds of the morning market. Vendors called to one another across the narrow streets. Children moved between the stalls with baskets. The smell of bread came from the bakeries on the eastern side. People gathered near the fountain at the center of the square.
Each is declarative and of comparable length, and each adds an item to the scene. The prose is not wrong; it is flat. No moment in it pulls the reader forward or holds them back. The scene does not breathe.
Consider it with variation introduced:
The morning market was loud before the sun cleared the rooftops. Vendors, children, the smell of bread from the eastern bakeries—and beneath all of it, the sound of the fountain, which the city had stopped hearing years ago.
The first sentence establishes pace. The second accelerates through its list and slows at the colon, bringing the reader to the fountain and to the quiet observation that follows. The scene has a center and a rhythm that leads toward it.
The Human Writer and the Spoken Word
A writer who reads their work aloud discovers things that silent revision conceals. The sentence that looked elegant on the page may stumble in the throat. The paragraph that seemed complete reveals that it ends on the wrong note.
This is not a sophisticated technique available only to poets. It is available to any writer willing to hear their own prose. The correction it requires is often simple: a word moved, a clause shortened, a sentence broken where it had been continuous. But the willingness to perform it depends on something a language model does not have: the understanding that prose is, at some level, always speech. It is interior speech, but it moves through the reader’s mind with the cadences of a voice. When they are missing, the reader feels it as friction—an effort to move through the text that the prose has not earned.
What a Human Author Does
A writer who has developed an ear does not think about rhythm as a separate consideration from meaning. The two are inseparable: the way a sentence sounds is part of what it means, and a sentence that sounds wrong is wrong, regardless of grammar and logic. The ear is not an ornament added to prose that is otherwise complete. It is the condition under which the prose becomes what it is trying to be.
This cannot be trained by exposure to text alone. It requires the body: the physical experience of carrying the sentence from its opening to its close without losing the reader on the way. A language model has read everything and heard nothing. The prose it produces is full of information and empty of sound, correct in its parts and deaf in its whole.
The reader does not finish a page of such prose and think: that was difficult to read. They set the book down and reach for another. The ear knows before the mind does.
Rio de Janeiro, the xxx day of June, MMXXVI.
