The Illusion of Authorial Voice: Why AI Text Feels Soulless
The Flawless Text No One Likes
To be honest, you wrote the prompt with great care: a clear audience, a warm but professional tone, short paragraphs, three examples, no clichés. The output is exactly that. The sentences are correct, the transitions smooth, the headings logical. And yet, after the third paragraph, something in you goes empty; it is as if the text belongs to no one.
This is precisely what we might call the illusion of the author’s voice: we assume that if the instruction is sufficiently precise, the author’s presence will be conveyed too. But an instruction produces a specification; voice comes from preference. The two are neither synonymous nor interchangeable.
Voice Is Not a List of Adjectives; It Is a History of Choices
When we say Ernest Hemingway has a recognizable voice, we are not pointing to a list of his adjectives. His short sentences mean nothing except alongside what he omits—what he himself called the “iceberg theory”: seven-eighths of every story is underwater, and only the tip is visible. Shortness of sentence in Hemingway is not a stylistic choice; it is a deliberate silence.
Virginia Woolf stands at the opposite pole; her long, wave-like sentences are not without reason either. She dissolves time in order to reveal a character’s mind. Neither of these is “better”; both are sets of repeated decisions that have cost something. Voice is exactly what a writer is willing to lose something for: readers, simplicity, or collective approval.
Voice is not in what the writer says; it is in what they are willing not to say, and in what they are willing to look bad for.
Why Prompt Precision Drives the Text Toward the Average
A language model produces the most probable continuation of the text. When you write “a warm and professional tone, with concrete examples,” the model moves toward the center of the distribution of all texts that satisfy this description. The result is a text that no one disagrees with and no one agrees with either. Three mechanisms are at work:
- Averaging: Every adjective-based description pushes the text toward the most common form of that adjective—that is, the safest version.
- Absence of cost: The model loses nothing; therefore it does not write a sentence that carries risk. Voice always begins with a small risk.
- Extreme balance: The “on the one hand… on the other hand” pattern recurs in outputs, because a neutral summary is always possible.
| A precise prompt determines | Voice is built from these |
|---|---|
| Audience, tone, length, structure | What the writer is unwilling to write |
| Roughly appropriate vocabulary | One unconventional word that only they choose |
| Relevant examples | Details that are not fully relevant to the topic and are real |
| A clear conclusion | A sentence that makes the reader abandon it midway |
Four Tests for Detecting a Voiceless Text
Before any editing, run the output through these four filters. If it fails one of them, the problem is not spelling; it is presence.
- The omission test: Which sentence would this hypothetical writer never have written? If you have no answer, that means no writer was present in the text.
- The position test: Is there a sentence in the text that someone could disagree with? A text everyone agrees with usually has said nothing.
- The rhythm test: Count sentence lengths. If they are all between 15 and 20 words, the text is not breathing; voice always has rhythmic irregularity.
- The strange-detail test: Anton Chekhov builds character with small objects: a torn galosh, a barking dog. Voiceless output has “example-like” details, not observed details.
Prompting for Voice: From Description to Examples and Prohibitions
The shorter route is not to give the model adjectives but examples. Put two or three paragraphs of your own writing in the prompt and ask it to take the syntactic pattern and level of detail from them, not their subject. Then add three constraints:
- Banned list: Explicitly prohibit the words and structures you are sensitive to (“in conclusion,” “it is worth noting,” “in today’s world”).
- Mandatory position: Ask the model to make and defend a specific, defensible claim rather than report both sides.
- Several distinct drafts: Ask for three versions with different positions; choosing among them is the moment your taste enters the text.
However, no prompt replaces human editing, because editing is the only stage at which someone takes responsibility for the text.
Human Editing: Where Voice Is Actually Installed
In the draft, instead of a wholesale rewrite, try these small moves: delete adverbs so the verbs can work; replace abstract nouns (“deep feeling,” “valuable experience”) with a concrete instance; deliberately break one sentence to disrupt the uniform rhythm; and in each section, add one “unnecessary” sentence that comes only from your own mind. It is precisely those seemingly redundant sentences that tell the reader someone is here.
Conclusion
AI is good at producing structured drafts and weak at producing preference. A precise prompt narrows the text’s range, and that very precision removes individual voice, because voice emerges from difficult choices, not from meeting specifications. If you want your text to have a soul, write the prompt to reach an editable starting point, and then decide for yourself what to keep and what to cross out. Voice is born at the very moment of crossing out.
