AI Writing That Sounds Like You: Rhythm Beats Tone Labels

TL;DR: AI writing that sounds like you is a data problem, not a prompting problem. Every model defaults to generic corporate phrasing because that is the statistical centre of its training. A tone label ("professional but warm") cannot pull it off that centre. A Writing DNA profile can, because it preserves the two things that make prose recognisably yours: your rhythm (sentence length, qualifier placement, transitions) and your word choice. Feed the profile to ChatGPT, Claude or Gemini and the model has evidence to follow instead of adjectives to guess at.

You asked ChatGPT for a quick note to a client. It returned "I hope this message finds you well. I wanted to reach out regarding our upcoming deliverable." You have never written either sentence in your life.

That gap between what you would have typed and what the model produced is the whole problem. Closing it takes something more specific than "make it sound like me."

Why Does AI Writing Default to Generic Corporate Phrasing?

Large language models are tuned to please the broadest possible audience, and the safest way to please everyone is to sound like no one in particular. The result is a house style shared across vendors: medium-length sentences, balanced structure, hedged claims, and a stock of polite openers and closers. That is what "generic corporate phrasing" means in practice, and it is the model's resting state.

We saw this directly when we measured five AI models across six style dimensions. Different vendors, different architectures, nearly identical output signatures. Sentence length converged, formality converged, and even the transition words converged. The models were not copying each other. They were all regressing to the same statistical mean because they were all optimised for the same thing: acceptable to the median reader.

A few markers show up in almost every unguided draft:

• Reflexive openers. "I hope this finds you well." "I wanted to reach out."
• Even sentence lengths. Fifteen to twenty words, over and over, with no short punch and no long run-on.
• Hedges stacked before the point. "It may be worth considering whether" instead of "Consider."
• Closers that cost nothing. "Please don't hesitate to reach out." "Looking forward to collaborating."

None of this is wrong. All of it is interchangeable. And that interchangeability is exactly what a reader detects when they say a message "sounds like AI." They are recognising the centre.

What Actually Makes Writing Sound Like You?

Writing sounds like you because of measurable habits, not because of a mood. The two that carry the most signal are rhythm and word choice, and both are far more specific than any tone label.

Rhythm is the shape of your sentences over time. Some people write in bursts: a twelve-word sentence, then four words, then twenty. Others hold a steady twenty-two. Rhythm also covers where you put qualifiers. One writer front-loads them ("Given the timeline, I think we should ship"). Another states the decision and qualifies afterwards ("Ship it. The timeline's tight, but it's the right call"). It covers your transitions too: whether you connect ideas with "so," "which means," "that said," or nothing at all.

Word choice is your working vocabulary under normal conditions. The verbs you reach for. The words you never use (many executives never write "circle back" in their lives, and the model reaches for it constantly). Whether you say "figure out" or "determine," "kick off" or "initiate," "quick note" or "brief update."

Here is the same message in a generic default and in a preserved voice:

The right column is not "friendlier" or "more casual" as a category. It has a specific rhythm (short, then medium, then short), a specific qualifier placement (decision first, reason second), and specific word choices ("pushing," "shout," "breaks"). Those are the things a profile has to preserve. A tone adjective preserves none of them.

This is also why the same person sounds different in Japanese, French or German without sounding like a different person. Register shifts with the language, as we cover in our guide to multilingual voice preservation, but the underlying rhythm habits and vocabulary preferences travel. Models flatten that distinction unless told otherwise.

Why Don't Tone Labels and Custom Instructions Fix It?

A tone label fails because it names a crowd, and a model given a crowd will pick the crowd's average. "Professional yet friendly" is true of thousands of people with completely different sentence rhythms. The model has no way to know which one you are, so it produces the median "professional yet friendly" writer, which is the corporate default with a smile added.

Custom instructions help more than nothing, and OpenAI's own custom instructions documentation is the right place to put voice rules once you have them. The failure is upstream of the feature. Most people fill that box with adjectives ("concise, warm, confident") because adjectives are what we have words for. Almost no one writes "average sentence length 14 words, decision before rationale, never open with a pleasantry, use 'figure out' not 'determine'." Yet that second version is the one the model can execute.

There is a second problem: AI does not know your tone. It knows the words that describe tone. When you write "warm," the model retrieves what warm-sounding text looks like across its training data, and that retrieval is, once again, the centre. You asked for a personal trait and received a category.

So the common fixes fail in predictable ways:

1. Adjective stacks ("professional, approachable, clear") describe a persona, so the model produces a persona.
2. "Write like this" with one pasted email gives the model a single data point. It imitates the surface of that one message and ignores everything the message did not happen to contain.
3. Editing the output afterwards works once, and then you do it again tomorrow and the day after. We wrote about why editing AI output is the slowest possible fix: the correction never makes it back into the instructions.
4. Vendor memory features remember facts about you, not your prosody. They know you work in finance. They do not know you never use semicolons.

Each of these is a prompt: a description of how you hope to sound. What is missing is evidence of how you actually write.

How Does a Writing DNA Profile Preserve Rhythm and Word Choice?

A Writing DNA profile turns your real writing into explicit, measurable rules the model can follow, so the generated content has somewhere to go other than the corporate default. That is the core of our position: the profile exists to preserve individual rhythm and word choice, and once those are preserved, generic phrasing has no gap left to fill.

The process runs in three stages, and the first is the one most people skip.

1. Collect samples that show your habits. Fifteen to twenty pieces of writing you actually sent, across the contexts you write in. Emails to clients, messages to your team, a longer document or two. The samples are the evidence. Our guide on how to select writing samples for AI analysis covers what to include and what to leave out (anything heavily edited by someone else, anything you wrote in a voice that was not yours).

2. Extract the patterns. This is where the profile earns its name. The analysis reads across your samples and records what is stable: your sentence-length distribution, your ratio of short to long sentences, where your qualifiers land, which transitions you favour, your recurring verbs, the vocabulary you avoid, how your register shifts between audiences. How style extraction works goes deeper into the dimensions, but the important point is that every one of them is stated as a rule with a value, not an adjective.

3. Deploy it where the model reads instructions. The output is a plain-text Style Profile. It goes into ChatGPT custom instructions or a Project, a Claude project's system prompt (Anthropic's system prompt guidance explains why instructions there are followed most consistently), or a Gemini Gem. Same profile, every tool.

The reason this outperforms a prompt is not magic. The model was always capable of writing at 14 words per sentence with the decision first and no pleasantries. It never did, because nothing told it to and its default pulled the other way. The profile replaces a guess with a specification.

Two consequences follow that are easy to miss.

Better models make the profile more valuable, not less. Model capability and knowledge of your writing are separate layers. Labs improve the first with every release and cannot supply the second. A model that follows nuanced instructions more faithfully follows your instructions more faithfully too, so each upgrade widens the gap between a profiled user and an unprofiled one.

The profile is an asset, not a session. A clever prompt lives in one chat window. A Writing Twin for ChatGPT, Claude and Gemini is a document you own, carry between tools, and update as your writing evolves. Switch vendors and it comes with you.

How Do You Test Whether AI Writing Sounds Like You?

The fastest test takes five minutes and needs no tools: generate a message, then check it against three of your own habits before you read it for meaning.

1. Count sentence lengths in the first paragraph. If every sentence sits between 15 and 20 words, the model is on its default. Your real writing almost certainly varies more than that.
2. Find the first qualifier. Is it before the decision or after? Compare with three emails you sent last week.
3. Scan for words you never use. "Delighted," "reach out," "align," "circle back," "moving forward." If two or more appear, the corporate centre is showing.

If the draft fails two of three, the fix is upstream, in the instructions. Adding "make it sound more like me" to the chat will move the output slightly and then it will drift back next session. Building a Writing Style Profile once, and pasting it where the model reads instructions, moves the default itself.

A more thorough version of this test: take one email you wrote, hide it, ask the model to write the same message with your profile loaded, then put both in front of a colleague who knows you. If they cannot pick yours reliably, the profile is doing its job. If they can, look at which of the three habits gave it away and check whether your samples actually showed that habit. Usually the gap is in the samples, not the model.

Where to Start

The order matters. Samples before rules, rules before prompts.

• Gather 15 to 20 pieces of writing you sent without heavy editing, across at least two audiences.
• Note three habits you can already see: typical sentence length, decision-first or context-first, and five words you would never write.
• Put those three rules in your custom instructions today. Even that small step moves the model off its centre.
• When you want the full pattern set, including the habits you cannot see in yourself, build a Writing Twin for ChatGPT, Claude and Gemini from your samples and deploy the same Style Profile everywhere you write.

AI writing that sounds like you was never about finding the right adjective. It is about giving the model evidence of your rhythm and your word choice, so the generic corporate default has nothing left to fill in.