Short answer: AI sounds like you when it keeps what makes your writing yours and drops the habits it adds by default. Know your own habits first (how specific you get, whether you tell stories, how strongly you take a side, how much you signpost, how you end), give the model your real details and examples rather than adjectives, cut the stock AI moves (openers, "first/additionally", lists of three, closing recaps), and correct drafts with specific feedback instead of regenerating. "Write casually" fails; "keep my story, take my side, no summary at the end" works. In WriteVoice you can train your double: it drafts in your voice and learns from every rating.
Why AI doesn't sound like you by default
Chat assistants are trained toward one helpful house style. Independent research on instruction-tuned models finds that style is noun-heavy and informationally dense, and that instruction tuning limits how well models imitate other styles (Reinhart et al., PNAS, 2025). Our own research finds that default AI answers are broad, signposted and wrapped up compared with people answering the same prompts, and that the details shift between model families and versions (see our research). So a vague request like "make it sound like me" gets pulled back toward the model's default.
Step 1: know what makes your voice yours
Adjectives like "friendly" or "professional" mean different things to everyone. Concrete habits don't. Our research on human and AI writing found five habits where people and chat assistants differ most in chat-style writing. They are the five dials in the WriteVoice quiz:
| Dial | What to look for | Where default AI tends to sit |
|---|---|---|
| Broad to Grounded | Dates, places, names and real numbers vs general truths | Broad: general statements over specifics |
| Impersonal to Personal | Your own experiences used as evidence | Impersonal: it rarely claims lived experience |
| Even-handed to Opinionated | A clear personal stance vs both-sides balance | Even-handed: "it depends", pros and cons |
| Direct to Signposted | Transitions, structure markers, lists of three | Signposted: "first", "additionally", triads |
| Open-ended to Wrapped-up | A closing recap or takeaway | Wrapped-up: a summary at the end |
The free WriteVoice quiz reads these five dials from about three minutes of chat (or pasted writing) and shows where you sit among people and where default AI sits on each. The dials where you are furthest from default AI are where an AI draft will sound least like you.
Step 2: give the model your material, not a description
- Give it the real details. The date, the place, the number, the name. Ask it to keep every one. AI drafts drift toward general statements when they have nothing specific to work with.
- Give it the story. If the point comes from something that happened to you, say what happened in a sentence. Otherwise the draft will make the point with "many people find".
- Say which side you're on. If you have a view, state it. Left alone, models balance.
- Show, don't describe. A short real example of how you write this kind of message beats "witty and warm".
- Remove anything private (names, addresses, health or money details) before you give it to any AI tool.
Step 3: cut the habits that give AI away
Most "this sounds like AI" reactions come from a handful of habits. Check a draft for:
- stock openers and sign-offs ("I hope this finds you well", "Feel free to reach out");
- signposting you wouldn't use ("First", "Additionally", "In summary");
- lists of three where you'd give one reason;
- a both-sides paragraph when you have an opinion;
- a closing recap or takeaway line;
- even sentence lengths, with no short ones;
- words you'd never say ("delve", "leverage", "thrilled").
More on what distinguishes AI writing, with data, in the tells of AI writing.
Step 4: correct, don't regenerate
When a draft is wrong, say what was wrong in concrete terms ("too many hedges", "sentences too even", "you used 'thrilled'", "you dropped my story"). Regenerating without feedback just samples the same default again. Specific corrections are what move a draft towards you.
The WriteVoice way: train your double
WriteVoice does this loop for you. After the quiz, open Train your double and give it something to write. It drafts it twice, the plain AI way and in your voice, and shows where each draft sits on your five dials. Rate the draft ("That's me", "Close", "Not me") and say what's off; your double learns from every rating, so the next draft is closer. Your double lives in WriteVoice: you write with it there.
Coming soon: connect your AI assistant to WriteVoice. We're building a connector that lets your own AI assistant ask WriteVoice to draft in your voice, using the double you trained. It isn't available yet.
Common mistakes
- Describing yourself instead of showing yourself. "Witty and warm" is how you'd like to sound; your real writing shows how you do sound.
- Mixing registers. Texts and reports are different voices. Most people write differently across contexts, which is why WriteVoice keeps a casual and a professional voice.
- Leaving out the specifics. A model can't keep a date or a story you didn't give it.
- Trusting "humanizer" tools. Rewriting to dodge an AI detector is a different goal from sounding like you, and detectors are unreliable anyway (see the tells of AI writing).
Frequently asked questions
Can ChatGPT learn my writing style?
Within limits. Chat assistants follow examples and specific feedback better than adjectives, but they drift back to their default style, and memory features store facts and preferences rather than the rhythm of your writing. WriteVoice's trained double is built to keep learning your voice from your ratings.
Why does AI writing still sound like AI after I give it examples?
Usually because of stock phrases, signposting, a closing recap and an even rhythm. Point out the specific habit in each draft, and give the model your real details so it has something specific to say.
What is the best way to make Claude write in my voice?
Give it your real details and a short example, and correct drafts with specific feedback. In WriteVoice, your trained double drafts in your voice; a connector that lets your own Claude ask WriteVoice to draft for you is coming soon.
Is it allowed to have AI write in my voice?
For your own messages, generally yes; check the rules of your school, employer or publication. Don't use voice-matching to impersonate someone else.