Does AI Know My Writing Style? Why Better Models Still Don't

TL;DR: AI does not know your writing style, and the next model release will not change that. Model capability and user-data conditioning are separate layers. Every lab improves the first and none of them can supply the second, because your writing was never in the model. Memory stores facts about you, custom instructions store adjectives, and adjectives describe thousands of people. The fix is to hand the model measurements of how you actually write. Better models execute those measurements more faithfully, which is why a real profile gets more valuable with every release, not less.

You have paid for ChatGPT Plus for a year. You have told it what you do, corrected its drafts hundreds of times, and it still opens every email with "I hope this finds you well." So you ask it directly: do you know my writing style? It says yes, and the next draft proves otherwise.

The honest answer is no. Nothing about how you write is in the model, and this post explains why that stays true no matter how good the models get.

Does AI Know My Writing Style at All?

No. A language model knows the statistical shape of the text it was trained on, plus whatever sits in the current conversation. Your writing was in neither, so when it drafts "as you", it is drafting as a blend of every writer who resembles the description you gave it.

That answer surprises people because the assistant behaves as if it knows you. Ask ChatGPT whether it has learned your style and it will produce a confident paragraph about your "clear, direct tone". The paragraph is generated the same way every other answer is generated: predicted from context. The assistant is not consulting a stored analysis of your writing, because no such analysis exists.

Three things get mistaken for style knowledge:

OpenAI's own memory documentation describes memory as details the assistant retains across chats, and the examples are factual: what you do, what you like, what you are working on. That is useful for context. It does not capture that your sentences average thirteen words, that you never hedge after a decision, or that you put the ask in the first line. We tested the gap in detail in our memory versus Style Profiles comparison.

Corrections vanish the same way. A rewrite you accept inside one chat shapes the rest of that chat, and only that chat. Open a new one and the model starts from its training distribution again.

Why Doesn't a Better Model Fix This?

Because model capability and user-data conditioning are separate layers, and a model release only touches the first one. Each generation gets better at reasoning, at following long instructions, at holding a register across a long draft. None of that adds a single sentence of yours to the weights.

Think of the two layers separately:

1. Capability layer. What the model can do with the instructions and text it is given. Owned by the lab. Improves with every release.
2. Conditioning layer. What the model is told about you, in this conversation, right now. Owned by you. Empty unless you fill it.

The labs work on layer one because that is the only layer they can reach. Anthropic does not have your board memos. OpenAI does not have your client emails. Even if a lab wanted to condition a model on you, it has nothing to condition it on. Fine-tuning on your writing is possible in principle and impractical for one person: it needs a large corpus, it locks you to one model version, and it has to be redone when the next release ships.

The practical consequence runs the other way from what people expect. A better model does not close the gap on its own. A better model widens the payoff of filling the conditioning layer properly, because it follows nuanced instructions more faithfully than its predecessor did. Hand a weak model a precise profile and it drops half the constraints by paragraph three. Hand a strong model the same profile and it holds them. The profile is the constant; the model is the variable that keeps improving in your favour.

We measured the baseline the models start from in our stylometric comparison of Claude, GPT and Gemini: each model has a house style of its own, and each one is a long way from any individual writer. Sampling knobs do not move it either, which is what our post on tuning your way out of average covers. Temperature changes variance, not voice.

What Does AI Actually Have When You Describe Your Tone?

Words. AI does not know your tone; it knows words that describe tone. When you write "professional yet friendly" into custom instructions, the model maps those words to the register it has learned for them, and that register is the same for every person who types the same phrase.

Run the experiment. Put "professional yet friendly, concise and warm" into a fresh chat and ask for a follow-up email after a client meeting. Ask a colleague to do the same with the same label. The two emails will resemble each other more than either resembles anything you sent last month. The label carried no information that separates you from your colleague, so the output cannot either.

That is the problem with every tone template on the web, and it is why "professional yet friendly" describes thousands of people. A voice lives somewhere the label cannot reach:

• Sentence length. A distribution, not a word like "concise". How short your short sentences get, how long the long ones run, when you switch.
• Qualifier placement. Whether "probably" opens the clause or closes it. Whether you hedge at all once you have decided.
• Transitions. "So" versus "therefore" versus a bare new sentence. Whether paragraphs open with the conclusion or build to it.
• Punctuation. Colon frequency. Comma density. Whether you use parentheses for asides or split them into their own sentences.

None of those are adjectives. All of them are countable, which is why they can be extracted from your writing and stated as instructions a model can follow. Our breakdown of how style extraction works goes through the dimensions we measure.

Why Do Pasted Samples Fall Short Too?

Samples beat labels, and they still leave the model guessing. Three emails show the model what you sounded like on three days. It has to infer which patterns are yours and which belong to that one message, and inference from a handful of examples drifts across a long conversation.

There is a second limit. Custom instruction fields have a size, and samples eat it. OpenAI's custom instructions guide gives you one field for how the assistant should respond. Fill it with three emails and there is no room left for the rules you actually need. Put the samples in a Project's knowledge files instead and the model reads them when it decides they are relevant, which is not every time.

The pattern across all three approaches is the same:

A prompt describes how you hope to sound. A profile built from your own writing is evidence of how you actually write. That is the whole difference, and it is why we compared the four instruction types side by side in our guide to the best ChatGPT custom instructions.

How Do You Give AI Your Actual Style?

Measure it once, state it explicitly, and paste the result into whichever assistant you use. The model does not need to learn you. It needs to be told, precisely, in a form it can execute.

This is what a Style Profile from MyWritingTwin.com contains: a Writing DNA analysis extracted from samples you already wrote, expressed as constraints. Sentence-length range and mean. Where your qualifiers sit. Which transitions you use and which you never use. Punctuation habits. Vocabulary you reach for and phrases you avoid. The output is a block of text you paste into ChatGPT custom instructions, a Claude Project, Gemini's saved info, or any other tool that accepts a system prompt. One profile, every model.

Two properties matter for the argument in this post.

The profile is a measurement, so it can be checked. When the model drifts, you can count sentence lengths in its draft against the profile and see where it missed. An adjective cannot be checked. "Was that friendly enough?" has no answer. "Was the mean sentence length within range?" does.

The profile is portable, so it survives model releases. Your Writing Twin for ChatGPT, Claude and Gemini does not depend on any one version. When a new model ships, you paste the same profile and get better execution of it. The asset appreciates with every release, which is the opposite of what happens to a template.

If you want to build the habit without the product first, do it manually:

1. Collect ten pieces you wrote and were happy with, in the register you need most (client email, internal update, proposal).
2. Count words per sentence across all ten and note the range and the typical value.
3. List every hedge you used and where it sat in the clause.
4. List the transitions you used and the ones you never used.
5. Write those findings as instructions, not adjectives, and paste them into your custom instructions.

That manual version is rougher than an extracted profile and still beats any tone label, because it gives the model numbers instead of words about numbers.

How Can You Test Whether Your AI Knows Your Style Today?

Give it a message you already sent, using only the brief you would have given a colleague, and compare the two drafts by counting. Five minutes, no tools.

Pick an email from last week. Write a two-line brief for it: who it is to, what you need, what they already know. Paste the brief into your assistant with your current custom instructions active and ask for the email. Then put its draft next to yours and count:

• Sentences per paragraph
• Words per sentence, shortest and longest
• Where the ask sits (first line or last)
• Hedges: how many, and whether they open or close the clause
• Commas per sentence
• Opening and closing lines

If the numbers differ, your AI does not know your style, whatever it told you when you asked it directly. If they match on one email, try three more in different registers. A tone label usually survives the first test and fails the second, because the label was written for one mood and you have several.

Run the same test after you paste a measured profile and the numbers converge. That convergence is the entire product, and it is also the reason better models help rather than hurt: a stronger model hits the numbers more reliably than a weaker one.

Nothing about you is in the model. It never will be. The good news is that it does not need to be, as long as you stop describing your voice and start measuring it.

You can see how far your current setup is from your real writing in about a minute with the Humanity Score test, or build a Style Profile from your own samples and paste it into the AI you already pay for.