AI and the Writer's Voice: What Can't Be Delegated
Consider a paragraph that has nothing wrong with it: no spelling errors, syntactically sound, with smooth transitions and a tidy conclusion. Now place that same text next to a note written by a specific person—a note that isn't necessarily more polished, but when you read it, you're certain someone is standing behind it and speaking. The gap between these two has become one of the serious questions for writers and editors: which part of writing does artificial intelligence take over, and which part can it never take over?
The short answer is this: language models learn "style," not "voice." But until the difference between style and voice is made clear, this sentence is nothing more than a pleasant-sounding slogan.
Style Can Be Copied, Voice Cannot
Style lives on the surface of the text: sentence length, vocabulary range, frequency of metaphor, preference for verbal sentences over nominal ones, and the degree of confidence the writer grants the reader. This layer is entirely describable. If you ask a model to "write in Hemingway's style," you'll get a result that carries the same markers: short sentences, stripped adjectives, unadorned dialogue. Hemingway himself explained this appearance in his iceberg theory; the visible part of the text is only the tip of the iceberg, and the other seven-eighths remains underwater.
This is precisely the point. The model reproduces the surface, but the decision of "what stays underwater" comes from the writer's experience. Its historical example is Gordon Lish's edits to Raymond Carver's stories—an intervention so heavy that the endings of certain stories changed, and the affair turned into a reputation-damaging dispute. Yet when you read Carver's stories, the voice is still his voice. Voice survives even heavy editing, provided it existed in the first place.
What Exactly Is a Writer's Voice?
Voice can be defined as follows: the writer's pattern of attention, plus their stance, plus the rhythm they bring to decision-making. People don't see everything the same way. Jalal Al-e Ahmad writes about himself and infertility in "A Stone upon a Grave," and the same anger he had in "Occidentosis" here takes on a private, unguarded form—something no model can extract from public data. Simin Daneshvar in "Savushun" looks at Zari with an empathetic yet precise distance. Joan Didion records details coldly and dispassionately to show how grief works.
Voice is the writer's pattern of attention, not their list of linguistic options. The model has the list and lacks the pattern.
Where Is Delegation Safe and Even Beneficial?
Before listing what is lost, let's be fair about what benefits. A large portion of writing work is mechanical, and delegating it is not only permissible but improves quality.
| Task | Delegation to AI | Reason |
|---|---|---|
| Gathering and summarizing sources | Fully | It's a precision task, not a taste-driven one |
| Suggesting structure and headings | Fully | It multiplies options; the choice remains yours |
| Removing filler and repetition | Fully | The text stays intact, it just gets leaner |
| Three alternative renderings of a paragraph | With caution | Useful for comparison; don't publish without choosing |
| Choosing sensory details and scenes | No | The model gravitates toward "average" details |
| Taking a stance, irony, endings | No | This is where the text finds its owner |
The dividing line is right here: precision tasks can be delegated, taste-driven tasks cannot.
Four Things Whose Delegation Leaves the Text Ownerless
- Lived detail. The model cannot distinguish real details from typical ones. "The sound of rain on the window" is nothing; "the sound of dripping on the cooler's tin" is something. The first came from nowhere, the second from a home.
- Judgment and risk. A text no one disagrees with is usually one no one remembers either. The model is structurally conservative and gravitates toward the safest sentence.
Five Tests to Tell Whether a Text Has a Voice
- The signature test: Remove the author’s name. If a regular reader cannot guess, there is probably no voice at work.
- The substitution test: Imagine this same text published under another name. If no one notices, you were merely an executor.
- The irreplaceable-sentence test: Find a sentence no one else could have written. If you cannot find one, the problem lies in the details.
- The omission test: Remove the final paragraph. If nothing is lost, the ending was not yours.
- The resistance test: Is there anything in the text that a reader might disagree with? A text everyone approves of has no stake of its own.
How It Works: AI as an Assistant, Not a Substitute
There is a simple rule when working with language models: let them suggest structure, but do not let them reconstruct your voice. Two practical approaches work.
First, the role of the “skeptical editor.” Do not ask the model to rewrite the paragraph; ask it to say what it did not understand, where the tone shifted, and where the argument remains unsupported. A damage report is safer than a repair.
Second, the 20/80 rule. Write the key parts of the text—the first paragraph, the last paragraph, and the turning point—yourself by hand, and move the middle sections forward with the model’s help. This small handwritten window connects the whole text to you. Remember that the model regresses toward the mean; if you rewrite five times in a row, the text becomes flawless, and perhaps nothing distinctive remains in it.
Conclusion
Delegating to AI is delegating time, not meaning. A good model can do in one working hour what used to take a day; but that same model cannot decide what is worth saying, from which angle it should be said, or where one should stay silent. A writer’s voice is made from these decisions, and these decisions are what convince the reader that someone is truly speaking to them. Keep the boundary clear: delegate meticulous work, keep taste-driven work for yourself. What remains is you.
