How to Make AI Write in Your Voice (2026)

You are reading Chapter 10 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

To make AI write in your voice, show it three things you already wrote instead of describing your style in adjectives, then run a five-point scan over the draft before it goes out. The capture takes about ten minutes, once. The scan takes about twenty seconds, every time after that.

This guide names the patterns that actually give AI writing away, walks through capturing your own voice from three pieces of your writing, and gives you the five checks to run before anything leaves your outbox with your name on it.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

The draft comes back grammatical, organised, and with nothing in it you could point at and call wrong. You read it again, something is off, and then you place it: it does not sound like you, and anyone who has read your emails for three years would notice inside two sentences.

That costs a senior professional more than it costs anyone else. A junior colleague sending an AI-flavoured email loses very little. A director sending one loses credibility that took years to build, because what your writing has been establishing all that time is that a specific person is behind it who knows what they think.

The usual fix makes it worse. Describing your style in adjectives, professional and approachable and concise, produces the exact default voice you were trying to get away from, because that description covers almost everything ever written at work.

One thing to settle before you start. This chapter asks you to paste your own writing into an assistant, so you need the rules you set in Chapter 9 about what can safely go in. Client names and live figures usually do not qualify.

Every assistant has a house style you did not choose. It reads as competent, polite and faintly corporate, and it is the most common reason a draft comes back unusable.

It is an average. The model was trained on an enormous quantity of published text, then tuned with human feedback to be helpful and inoffensive to as many readers as possible. What comes out is the middle of that distribution, the register that would pass in the largest number of situations. Acceptable to everyone reads, in practice, as written for nobody.

So the default is not an absence of style. It is a style, with preferences, and you get its style instead of yours every time you ask for a draft without saying anything about voice.

You can see it in the sentences. Even, medium-length lines all the way down. The formal word where a plain one would do. Balanced on every side of every question. Nothing in it is wrong, and nothing in it is only yours.

The fix is that you already have years of your own writing sitting in a sent folder. The rest of this chapter is about what to do with three pieces of it, and what to look for in the twenty seconds before you press send.

Most of what circulates about spotting AI writing is folklore, and some of it is backwards. What gives a draft away is how the sentences are built, and how many of them are built the same way.

Start with the famous one, because it is wrong. In August 2026 The Economist compared 1.2 million words of its own journalism against versions of the same articles written by ChatGPT, Claude, Gemini and Grok. Of the four models, only Claude used em dashes more often than human writers did, and ChatGPT used noticeably fewer than everyone. If you have been stripping dashes out of your writing to avoid suspicion, you have been editing for a rumour.

The same comparison found the punctuation signal runs the other way. Models use fewer commas and semicolons than people do, and almost no parentheses. Their single most overused word was "and".

Sentences run long, and they run long consistently, with paragraphs rarely broken by a short statement. They reach for the polysyllable over the plain word: significant, increasingly, consequences, and scientific vocabulary like parameter and methodology in places where something shorter was available.

Then the scaffolding. Not X, but Y. Not only X, but also Y. The list of exactly three where the third item was added to finish the shape. Good writers have used all of these for centuries, which is why hunting them one at a time is useless. The finding that matters is rate: the models reached for them more often per thousand sentences than the humans they were set against.

The last family is the one colleagues feel without being able to name. The draft has no view in it. Every side gets its due, everything is worth noting, and claims are attributed to "industry reports" or to "experts" who are never identified. For a senior person that is the expensive part, because taking a position is most of what people read you for.

Density is the real test, not presence. Good writers use every device on this page. Finding five of them together on one page is the tell.

You will want to settle it with a detector. Do not, and be wary of anyone who runs one on you. OpenAI withdrew its own AI text classifier in July 2023, saying it caught 26% of AI-written text while flagging 9% of human writing as machine-made. A Stanford study the same year found popular detectors misclassified more than 61% of essays by non-native English speakers as AI-generated. Treat a detector score as an opinion, never as evidence, in either direction.

Rewriting a draft that already reads this way is its own job, and we have a full guide to humanizing AI writing that covers it. This chapter stays on the two things that sit either side of that: capturing your voice once, and the twenty-second check.

Describing your style in words works badly and takes longer than the alternative. Showing it three pieces of your own writing works immediately, and the capture only has to happen once.

Same format and same register as the thing you want drafted. Three sent emails if you want an email, three posts if you want a post. A few hundred words each is plenty, and choose ones you were pleased with rather than ones that happened to be long. Do not mix formats: a board memo and a LinkedIn post average into something that is neither.

Do not ask it to imitate anything yet. Ask what it notices about how you write: how long your sentences run, which words you reach for, how you open, how you close, what you never do. What comes back is a description of your own writing, and it will be more useful than anything you would have written about yourself, because nobody is a reliable witness to their own habits.

It will get two or three things wrong. It will also state as a rule something you did once by accident, so delete that, and add the thing you know about your writing that it missed. This is the step people skip, and it is what turns a generic description into something worth reusing for a year.

What you are left with is a short paragraph that describes how you actually write. That paragraph is the asset. Call it your voice note.

From then on, drafting starts with the voice note pasted above the request, plus one sample when the piece matters. It costs five seconds and it is the difference between editing a draft and rewriting one. Examples are a prompting technique in their own right, well past the question of voice, and the 26 principles of prompt engineering covers the wider set.

Be realistic about the ceiling while you are at it. Samples get you most of the way and not all of the way, and the rest is you, editing. That is why the next section is a scan you run before anything leaves your outbox.

You will not catch this by reading carefully. You catch it by reading fast, once, hunting for five specific things, the way you scan a contract for the number instead of reading the clauses.

Run your eye down the right-hand edge of the paragraph. If every line ends in roughly the same place, cut one long sentence in half and let the short half stand on its own. Like that.

Search the draft for "not" and see whether you have written "not X, but Y" twice on one page. Count any list of three and ask whether the third item is doing work or completing a shape. Keep one, delete the rest.

Utilise. Facilitate. Robust. Find the word you would never use in a meeting and swap it for the one you would.

If every paragraph is even-handed, write the sentence that says what you think. This check and the next one are about substance rather than style, and they are the two that matter most.

Detail is what an assistant cannot supply, because it was not there. A real name, a real number, a real date, what was actually said in the meeting on Tuesday. Add the specific one.

Two habits are worth keeping past this chapter. Read the first line aloud every time, because the opening is where the default voice shows itself most plainly. And treat yourself as editing rather than approving: a draft you changed five things in is a draft you wrote. Four or five small edits is the normal result, and it is enough.

The chapter closes with three exercises that take about fifteen minutes. Name the three samples you will use, write your corrected voice note, and run the scan on a draft you already have.

Before: you ask for a draft, it comes back clean and slightly off, and you either send it as it is or rewrite it from the first line.

After: you paste a voice note you wrote once, the draft comes back closer to your own writing, and you spend twenty seconds and four or five small edits on it. You can also say what gave the old drafts away, which means you can see it coming.

Two things carry into Chapter 11: your corrected voice note, and one draft you have already run the scan on. Chapter 11 is about where instructions like that live, in saved instructions, projects and custom setups, so a new conversation does not begin from zero. If you want a one-page reference to keep beside you while you work, start with our free AI cheat sheets, and the rest of the syllabus lives in the AI Central Library.

Show it three things you already wrote in the same format as the thing you want drafted, ask what it notices about how you write, correct the description it gives you, and keep that corrected paragraph as your voice note. Paste the voice note above every drafting request from then on. Describing your style in adjectives does not work, because it produces the default voice you were trying to escape.

Mostly no. In August 2026 The Economist compared 1.2 million words of its own journalism against versions written by ChatGPT, Claude, Gemini and Grok, and of the four models only Claude used em dashes more often than human writers did. ChatGPT used noticeably fewer than everyone.

Construction rather than characters. Long sentences of uniform length, the formal word where a plain one would do, "not X, but Y" scaffolding and tidy lists of three, fewer commas and semicolons than people use, and a draft that takes no position and names nobody. Any one of these is normal in good writing. Five of them on one page is the tell.

No. OpenAI withdrew its own AI text classifier in July 2023, saying it caught 26% of AI-written text while flagging 9% of human writing as machine-made, and a Stanford study the same year found popular detectors misclassified more than 61% of essays by non-native English speakers as AI-generated. Treat a score as an opinion, never as evidence.

Three, in the same format and register as the thing you want drafted, a few hundred words each. Pick ones you were pleased with rather than ones that happened to be long, and do not mix formats, because a board memo and a LinkedIn post average into something that is neither.