In 2023, literary agents reported receiving floods of AI-assisted query letters — hundreds of proposals for books on identical trending topics, using nearly identical phrasing. Most were rejected within seconds. One agent, Janet Reid of New Leaf Literary, noted publicly that the difference between the slush pile and a request for pages was almost always a single factor: whose specific perspective made the idea irreplaceable.
Voice isn't vocabulary. It's the accumulation of choices — what you notice, what you skip, how fast you move, when you're funny, when you're still — that makes a reader feel they are with a particular human being.
Most writers think of voice as style — word choice, sentence length, whether you use contractions. That's the surface layer and it's the easiest for AI to approximate. Below it are two deeper layers that AI cannot replicate because they require lived experience.
Diction, sentence rhythm, punctuation habits, formality level, use of humor, paragraph length, transition words. These patterns can be learned from examples and reproduced statistically.
The specific things you've noticed, the opinions you actually hold, the comparisons only you would make, the gaps you leave because you know the reader already knows — these come from living a particular life.
Voice researchers and writing teachers have studied this for decades. The short answer: voice is recognized by specificity and omission. What the writer chooses to include — and what they trust themselves to leave out — reveals perspective more than any stylistic flourish.
Consider two descriptions of the same event:
"The concert was amazing. The crowd was excited. The music was loud and the performance was energetic."
"The bass hit somewhere in my sternum. The woman next to me had been crying since the opening chord and she didn't care who saw. I've been to maybe forty concerts. This was the one I'll describe to my grandchildren."
The second version assumes things: you know what a sternum is, you've had the "this is the one" feeling before. It omits things: it never says the genre, the venue, the artist's name. That trust — in both writer and reader — is voice.
AI language models are trained on the statistical center of human writing. Their outputs trend toward the most commonly co-occurring word sequences — which means toward averaged, conventional expression. When you ask an AI to "write in my voice" without guidance, it produces the voice of Everyone, which is the voice of No One.
This is not a flaw to be fixed. It is the architectural reality of how these models work. Understanding it lets you use AI precisely where it helps (generating raw material, varying structure, filling in facts) while keeping yourself in the seat where only you belong: deciding what the piece is actually saying.
AI can write words. Only you can write your words. The distinction is not about effort — it's about origin. Your specific perspective, formed by your specific life, is the only resource in this workflow that is genuinely non-renewable.
Before you can protect your voice in AI-assisted writing, you need to know what it actually consists of. In this lab, you'll work with an AI coach to build a personal voice inventory — the specific markers that make your writing yours.
The AI will ask you questions about your writing habits, preferences, and tendencies. Answer honestly and specifically. After 3+ exchanges, you'll have a working voice profile to use in later labs.
In February 2023, Clarkesworld Magazine — one of science fiction's most prestigious short fiction venues — temporarily closed submissions after the volume of AI-generated stories overwhelmed its slush pile. Editor Neil Clarke reported that the number of spam submissions (his term for AI-generated stories without disclosed AI use) had risen from single digits monthly to hundreds within weeks.
What made the AI stories identifiable wasn't that they were bad in the conventional sense. They were competent. They were readable. They had beginnings, middles, and ends. What they lacked was any sense that a particular person needed to write this particular story. They felt, Clarke noted, like stories produced by someone who had read about stories rather than lived through experiences that demanded expression.
AI doesn't just remove bad writing. It removes the eccentric writing — the unexpected choices, the jarring turns, the moments where a writer's specific sensibility overrides conventional technique. Here are five specific mechanisms through which this happens:
The most insidious effect of AI on voice is not that it makes writing worse. It makes writing professionally adequate — which, for most creative and persuasive purposes, is worse than bad. Bad writing at least signals a human trying. Adequate writing is forgettable.
In 2023, researchers at the University of Pennsylvania's Wharton School found that when business writing was AI-edited, evaluators consistently rated it as clearer but less memorable and less persuasive. The editing improved surface clarity while removing the distinctive assertions that made readers remember and act on the content.
In the Wharton study, AI-edited memos scored 23% higher on "clarity" ratings but 31% lower on "I would act on this" ratings compared to original human drafts. Clarity and persuasiveness moved in opposite directions.
When AI flattens your voice, the specific losses are: the reader's sense of your credibility (because your unusual observations prove you've actually thought about the subject), the emotional investment that comes from feeling a real person wrote this, and the memorability that comes from encountering something you hadn't seen that way before.
None of these can be added back by further AI editing. They have to be preserved during the editing process — which means knowing, precisely, what you're protecting before you start.
In this lab, you'll paste a short piece of your own writing (3–6 sentences) and an AI-edited version of it. The AI coach will identify exactly which of the five flattening mechanisms occurred and what was lost.
If you don't have an AI-edited version, paste just your original writing and ask the AI to "edit it to improve clarity" — then analyze what changed.
In September 2023, George R.R. Martin joined a class-action lawsuit against OpenAI, alleging his distinctive voice and narrative style had been used without permission to train language models. The legal argument aside, the case illuminated something technically significant: Martin's voice was recognizable enough to be identified, documented, and claimed as property.
The characteristics cited were specific: his use of dramatic chapter-ending revelations, his management of point-of-view across ensemble casts, his particular mixture of high-register archaism and colloquial brutality. These weren't vague stylistic impressions — they were enumerable, teachable elements. If you can enumerate your voice, you can protect it.
Voice protection in AI-assisted editing requires a three-stage process. Each stage addresses a different vulnerability window.
The most effective prompt structure for voice-preserving AI editing has three components:
1. Role definition: "You are a copy editor, not a ghostwriter."
2. Constraint list: "Do not change sentence rhythms, do not soften claims, do not remove personal observations."
3. Specific task: "Fix only: subject-verb agreement errors, inconsistent tense, and unclear pronoun references."
This is dramatically more effective than "edit this for me" — which gives AI the latitude to apply all five flattening mechanisms. The more specific your constraint list, the more your voice survives the edit.
A highly effective technique: ask AI to produce three alternatives to a specific sentence rather than a complete rewrite of a paragraph. Then compare all three to your original and cherry-pick — or use the alternatives to spark your own revision. You remain the decision-maker; AI generates options.
This approach was documented in a 2023 study by Stanford's Human-Computer Interaction group, which found that writers who used AI to generate options rather than replacements reported significantly higher satisfaction with final drafts and rated the work as more authentically theirs.
Before asking AI to make any change, ask it to explain what it would change and why. This creates a review layer where you can accept the reasoning or reject it before the text is altered. It also trains you to recognize which of your habits are genuine voice choices versus genuine errors — a distinction that's harder to see in your own work.
If an AI edit makes your writing sound like it could have been written by anyone with your rough topic knowledge, it has removed your voice. If it makes your writing sound more precisely like the thing you were trying to say, it has helped. The test is always: whose thought is this now?
In this lab, you'll practice the full MARK / LOCK / RESTORE workflow with a real piece of writing. Bring 1–3 paragraphs you've written. The AI will help you build a voice-preserving prompt, execute a constrained edit, and then identify what to restore.
The goal is not to get a perfect draft — it's to practice making the decisions that keep the work yours.
In late 2023, The New York Times released internal guidelines for AI-assisted journalism. One of the central requirements was that any AI-generated text must be substantially rewritten by the journalist before publication — not merely fact-checked, but transformed through the reporter's specific knowledge, sources, and perspective.
The Times editorial leadership framed this not as a restriction but as a definition: the journalist's job is not to verify AI output — it is to replace AI's generic framing with their own reporting and judgment. The AI draft is raw material, not a finished product requiring polish. The journalist's voice isn't added on top; it is the product itself.
When working from an AI-generated draft, your voice has to be actively injected at four specific points. These are the sites where AI output is most generic and where your specificity has the highest impact:
A practical benchmark used by professional writers working with AI: if fewer than 40% of the sentences in a final draft originated with you — either written by you originally or substantially transformed by you from AI output — the piece probably doesn't have your voice in it. It has your name on it.
This isn't a moral rule, it's a quality rule. The 40% threshold is roughly the minimum needed for your specific knowledge, opinions, and examples to dominate the reader's experience. Below it, the AI's generic framing tends to win.
Having your name on a piece and having your voice in it are different conditions. Your name is a byline. Your voice is the evidence that you — specifically — thought about this subject and found something particular to say. Readers can tell the difference, even when they can't articulate why.
Professional writing coaches who work with AI-assisted drafts use a specific exercise: take an AI-generated paragraph and rewrite every sentence in it until you'd be comfortable defending each one in a conversation. "Defending" doesn't mean proving — it means that each sentence reflects your actual understanding, not just a plausible-sounding claim.
This exercise, practiced regularly, builds what voice coaches call sentence-level ownership — the habit of not publishing a sentence you couldn't explain in your own words from your own knowledge base.
AI can build a scaffolding. You have to live in it. The scaffolding tells you where the rooms are; your voice tells the reader what it's like to be there. No amount of AI generation can substitute for a writer who has genuinely inhabited their subject — and the reader always knows the difference between a building with someone in it and an empty structure with lights left on.
In this lab, you'll work through the four injection points with an actual piece of writing. Either bring an AI-generated draft you want to transform, or ask the AI to generate a generic draft on a topic you know well — then work through replacing the opening, the hedges, the examples, and the closing with your own voice.
The AI coach will help you identify each injection point and evaluate whether your replacement is genuinely yours or still generic.