Why the Default Voice Costs You
Ask a model for an essay and it hands you a specific person: relentlessly balanced, fond of "moreover" and "in today's fast-paced world," allergic to a strong opinion, addicted to the tricolon ("clear, concise, and compelling"). That person doesn't exist, which is exactly the problem. Graders read forty essays a weekend. By essay six they can smell the default voice — the hedging, the throat-clearing intros, the conclusion that restates the intro. Readers feel it too, even when they can't name it.
The default voice isn't bad writing. It's averaged writing. The model is predicting the most probable next word across millions of documents, so it lands on the blandest version of every sentence. Your job in this chapter is to drag it off that average and onto your specific way of saying things.
Give It a Style Sample, Not Adjectives
Most people try to fix voice with adjectives: "write in a casual, engaging tone." This barely works. "Casual" to the model is a vague cloud that still resolves to the default. The model doesn't know your casual.
Show, don't tell. Paste 200-400 words you actually wrote — an old essay, a text thread, a journal entry — and make the model study it before writing anything.
Here are three paragraphs I wrote. Study the voice: sentence
length, rhythm, how formal it is, what kinds of words I avoid.
[paste your writing here]
Now in two sentences, describe my writing style back to me.
Don't write anything else yet.
Making it describe your style first does two things. It forces the model to actually parse the sample instead of skimming it, and it gives you a checkpoint — if its description is wrong, your sample was too short or too inconsistent. Once it nails the description, then ask for the draft "in that exact voice."
If you don't have a clean sample, write five sentences right now about anything — your weekend, a movie you hated. Five honest sentences beat zero. The model needs a target, and any real text outperforms a list of adjectives.
Build a Reusable Voice Guide
A style sample fixes one session. A voice guide fixes all of them. Spend ten minutes writing a short spec you paste at the top of every writing prompt:
MY VOICE GUIDE
- Sentence length: mostly short. One long sentence per paragraph, max.
- Tone: direct, a little skeptical. I'd rather be blunt than polite.
- Banned words: delve, moreover, furthermore, tapestry, "in today's world",
"it's important to note", "navigate the landscape".
- No rhetorical questions as openers.
- I use "I think" and "honestly," not "one might argue."
- Contractions always. "Don't," not "do not."
- End sections on a flat statement, never a summary.
The banned-words list is the highest-leverage line in the whole guide. The AI sound lives in a small vocabulary of tells — delve, tapestry, testament, realm, moreover, crucial, vibrant, in conclusion. Ban them by name and you delete most of the giveaway in one stroke. Build your list by reading any AI draft and circling every word a real student wouldn't say out loud.
Keep the guide in a note on your phone. Paste it before every request. It's the difference between re-explaining your voice forty times and doing it once.
Edit Toward Yourself, Not Away From Errors
Even with a good guide, the first draft drifts back toward the default — longer sentences, safer claims, a creeping "moreover." Don't accept it. Run a second pass aimed only at voice:
This draft sounds too much like generic AI writing. Rewrite it so it
sounds like a sharp 20-year-old wrote it, not a marketing department.
Cut every sentence that hedges. Delete any word from my banned list.
Break up sentences longer than 20 words. Keep my argument identical.
Then read it out loud. This is the single best detector you own. If a sentence is hard to say or sounds like a brochure, it's not yours — cut it or rewrite it by hand. You don't have to accept the model's wording. The strongest move is to take its draft and retype one paragraph entirely in your own words. That one human paragraph re-anchors your voice and makes the rest easier to fix.
Watch for the structural tells too, not just words: paragraphs that all start the same way, every list having exactly three items, conclusions that announce themselves. Real writing is lumpy. Vary your openings, let a list have two items or five, and end where the thought ends.
If you want to go deeper on steering models with precise instructions, the Advanced Prompt Engineering course covers techniques that transfer directly to voice control. And because the default voice is partly a misinformation risk — it states weak claims with fake confidence — the habits in AI Literacy: Spot Misinformation pair well with sounding like yourself.
The Test That Matters
Before you submit anything, run one check: would a friend who knows you believe you wrote this? If the answer is "it sounds fine but it doesn't sound like me," you haven't finished. Fine is the default voice's whole brand. Yours should sound like a specific person with specific opinions — because that's the only voice a grader remembers and a reader trusts.