How You Think Is the Missing Layer in Your AI Writing Profile

I fed the tool my own writing first.

Ten samples. Real emails, a couple of posts, one memo. The profile came back fast. On the surface, it was me. The vocabulary was right. The sentence rhythm was close. The formality landed where I set it.

Then I asked an AI to draft with it. The draft sounded like me. But it argued like someone else.

It opened with the conclusion. I never do that. I lead with the evidence and let the reader arrive at the point on their own. It hedged where I would have committed. It reached for a sports metaphor. I have never once reached for a sports metaphor.

The words were mine. The thinking was not.

That gap is the whole story.

Samples show the output, not the process

A writing sample is a finished thing. It is thinking, cleaned up and sent.

By the time you read one of my emails, the argument is already built. You see the bricks. You do not see how I laid them. You do not see the three orders I tried first. You do not see where I hesitated, or why I cut the qualifier at the end.

Style analysis is very good at the bricks. Vocabulary, cadence, punctuation, formality. Fifty-plus markers, and they hold up across tools. That part works.

But style is the surface. Underneath it there is a structure. How you build a case. Whether you hedge or commit. Whether you explain the why or just state the what. Which worlds you borrow your metaphors from.

That structure does not show up cleanly in polished writing. Polishing is the act of hiding it. It shows up when you think out loud.

Where you already think out loud

You have a record of yourself thinking out loud. You just do not file it that way.

It is your AI chat history.

Think about how you talk to a chatbot. Do you polish it first? Of course not. Every prompt you write is raw cognition. You are not performing. You are working. You correct the model when it is wrong. You push back on a weak answer and say exactly why. You ask the question behind the question. You reframe when the first framing fails.

The Cognitive Layer

The part of your profile built from imported AI chat history rather than writing samples. It captures how you reason: your argument direction, your certainty, your habit of explaining causes, the domains you draw analogies from, and whether you break problems apart or explore them whole.

I exported mine. Two years of it. ChatGPT, Claude, and Gemini. Around two thousand conversations. Close to eight thousand of my own messages. Fifty-odd thousand words of me, thinking.

Then we ran it through analysis pointed at a different question. Not how does he write. How does he think.

Five things it reads for

The cognitive layer looks for patterns you can measure. Five of them.

Argument direction. Do you lead with evidence and build to the point, or state the point and defend it? I build. Other people declare. Both work. They read completely differently.

Certainty. Do you hedge or commit? "This might be worth considering" is a different person than "This is the issue." The model has to know which one you are before it speaks for you.

Causal reasoning. Do you explain why, or just state what? Some people always give the reason. Others trust the reader to ask. Get this wrong and the draft either over-explains or reads as curt.

Analogy domains. Which worlds do you reach into? I reach for operations and mechanics. Not biology. Not sports. The domain you pick is a fingerprint.

Analytical or intuitive. Do you take a problem apart, or feel your way through it whole? Most of us do both. We switch at predictable moments. That switch is learnable.

None of this is about vocabulary. It is about the shape of the thought before the words arrive.

Picture a single request handled two ways. Ask for a project update. The style-only draft opens with the verdict: "We are two weeks behind." Clean, correct, and not how I would ever say it. The version that knows how I think opens with what happened, walks through why, and lands on the two weeks at the end, where I would put it. Same facts. Same vocabulary. A different mind behind them.

Why chat history, and not more samples

We made a deliberate call. Cognitive patterns come only from imported chat history. We do not guess them from your writing samples.

We could. It would be easier. It would also be a guess dressed up as data. If we inferred your reasoning from your finished writing, we would be inventing a person and calling it you.

I would rather show you less and have it be true.

The science behind it

I did not want to build this on a nice-sounding idea. So we went to the research first.

In 2025, a team led by Vasudha Varadarajan at H. Andrew Schwartz's lab published the result that convinced me. They had people describe a recent decision in their own words. Then they ran those same people through a classic decision-making experiment. The question was simple. Could the language alone predict how each person actually reasoned?

It could. Around 0.8 AUC. Not perfect. Far better than chance, and good enough to act on. The signal was not in the fancy words. It was in the shape of the thinking. You can read the paper here: Capturing Human Cognitive Styles with Language, NAACL 2025.

There is a deeper reason it works. The small words you never think about, the "because," the "but," the "maybe," get used below the level of conscious choice. James Pennebaker spent decades showing that these throwaway words track how a person reasons, hedges, and explains. Whole fields run on this now. LIWC scores your "analytic" thinking and your causal language straight from word counts.

That is the part I find almost unfair. Your style, you can perform. Your cognition leaks. It shows up in the joints of your sentences, and it holds steady over time. You can swap your vocabulary in an afternoon. You cannot easily change whether you argue from evidence or from conclusions. The stable thing is the thing worth capturing.

What is live, and what is still coming

Here is the honest state of it.

Live today: the whole supervised import. You consent before anything is read. You upload your ChatGPT, Claude, or Gemini export. We parse it, pull out your side of every conversation, and extract the five cognitive dimensions. Then we drop the raw file and hand you the receipt. It shows the timestamp your file was gone, with only the patterns kept. Nothing reaches your profile until you see a preview and approve it. We validated the extraction on real data, mine included, against style-only output. The version that knew how I think won. It was not close.

Still coming: correcting the read one line at a time. Right now you approve the whole preview or you skip it. Soon you will be able to push back on a single insight. If the tool says you rarely hedge and you were just venting that day, you tell it so, and the rest stands.

That is why we built consent and the deletion receipt first. Trust before speed. The fine-grained control is the last piece.

I will not pretend every corner is finished. The engine works. The privacy flow works. The per-insight editing is what is still going up.

The part that surprised me

I expected the cognitive layer to make the output more accurate. It did.

I did not expect it to change how I saw myself.

Reading my own patterns back was strange. I lead with evidence, always. I almost never hedge. I explain the why compulsively, even when nobody asked. I had felt these things for years. I had never seen them scored.

Your writing is how you sound. Your thinking is who you are underneath the sound. For a long time, AI could copy the first and miss the second. That is why the drafts always needed a rewrite. They had your voice and someone else's mind.

We are closing that gap.

You do not need to teach AI to write like you. You have done enough of that. Teach it to think like you first. The writing follows.

Build your AI Writing Profile and start with the layer that already knows how you write. The one that learns how you think is on its way.