The science of sounding like yourself

Feed the same prompt to the same model on ten different afternoons and you get back ten drafts that read as though one mid-level consultant wrote all of them on a deadline. Competent. Clean. The kind of prose that never embarrasses you and never quite surprises you either. That flatness is not a failure of your prompt or your imagination; it is a measurable property of how the model was built, and the quietly hopeful thing about a measurable property is that you can measure it back out.

So let us look at what the machine is actually doing, because once you can name the sameness you can start taking it apart. The research that explains it is older than the chatbots and older than the web. It came from people who spent whole careers proving who wrote what.

01Your fingerprints are made of small words

Imagine someone hands you two essays with the names torn off and asks which of two known authors wrote each. You would reach for the big content words, the nouns and the arguments and the subject matter. You would be wrong to. The signal that actually gives a writer away lives in the words nobody notices: the, of, that, while, upon, by. Function words, the grammatical glue you stopped seeing somewhere around the second grade.

In 1963 two statisticians, Frederick Mosteller and David Wallace, used exactly that idea to settle one of the oldest arguments in American letters. Twelve of the Federalist Papers had been claimed by both Alexander Hamilton and James Madison, and historians had spent a century and a half failing to agree. Mosteller and Wallace ignored the politics and counted the filler. Madison wrote whilst; Hamilton wrote while. Madison reached for upon far less often. Tiny unconscious habits, stable across every subject the two men ever touched, and they were enough to assign all twelve disputed papers with a confidence that has held up ever since.

That is the founding insight of stylometry, the quantitative study of style, and it has only sharpened in the sixty years since. The words you pick when you are not paying attention, the ratio of pronouns to prepositions, the distance you tend to run a sentence before you let it stop, all of it forms a signature about as individual as handwriting. You did not design it, and you mostly cannot feel it. It is yours anyway.

02Two dials that decide how you sound

Counting function words tells you that two writers differ. It does not tell you how. For that you want the framework a linguist named Douglas Biber built in the late eighties, when he ran a then-enormous pile of texts through a factor analysis and asked which features liked to show up together.

What fell out were dials, not boxes. The sturdiest of them runs between what Biber called involved and informational production. Involved writing leans on the reader: contractions, present tense, the word you, little verbs of thinking like I suspect and I doubt. It sounds like a person talking. Informational writing leans on the subject instead: dense noun phrases, careful abstraction, the speaker quietly stepping out of frame. It sounds like someone who has edited themselves twice. Neither is the right answer. Most of us live at a fairly fixed spot on that line and barely move from it across a lifetime of writing.

A second dial runs between narrative and analytic, between recounting what happened and dissecting what is true. You can probably feel your own default already. Some writers reach instinctively for a scene, for somebody doing something on an ordinary Tuesday. Others reach straight for the principle underneath. Biber found sixty-seven of these features clustering into a handful of dimensions, and the useful part for our purposes is that a large language model, having swallowed most of the linguistics literature during training, already knows what every one of them means. You can simply hand it the coordinates.

Your style is measurable, and the machine already speaks the language of the measurement.

03The tells

If the model has a fingerprint of its own, and it does, that fingerprint is built from habits it cannot stop performing. You have started to spot them. There is the vocabulary that appears nowhere in your writing and everywhere in the machine's, the delve and the tapestry and the eternal rich tapestry. (This sentence has now used all three, under protest, and intends never to again.)

Past the vocabulary, the architecture gives it away faster. There is the rule of three, the reflex to bundle everything into a tidy trio whether or not the third item has anywhere to be. There is the connective scaffolding, a Moreover here and a Furthermore there, as if the logic could not hold its own weight. The fake-profound inversion turns up constantly, that it-is-not-merely-X-but-Y construction that borrows the cadence of insight and smuggles in nothing. And then the symmetry, paragraphs swelling to identical size and tying off on identical little bows, coupled together like train cars.

None of this is the model being stupid. It is the model being average, on purpose. Trained to satisfy the widest possible audience and then polished by human raters who reward the safe and the fluent, it slides toward the center of everything it has ever read. The center is where prose goes to become forgettable.

04Humans are lumpy

The tells all circle the same buried fact. Real writing is uneven. You build a long, accumulating sentence that gathers its qualifications as it goes and lands somewhere it has earned, and then you stop short. One beat. You change your mind in the middle of a paragraph. You let a flat transition stand because the logic was already plain and narrating it would have insulted the reader. Measured across a page, your sentence lengths jump around far more than a model's do, and that jumpiness, which the field calls burstiness, is one of the steadiest signals that a person rather than a machine was at the keyboard.

There is a measurement for the same idea coming from the other direction. Perplexity asks how surprised a language model is by each next word in a piece of text. Machine prose scores low, which makes sense, since it is choosing the least surprising word nearly every time. Human prose scores higher, because you keep reaching for the odd exact word, the one the statistics did not see coming. Surprise, in the right dose, reads as a mind.

05What a voiceprint actually is

All of which lands us, finally, on the practical question. If your style is measurable, and the model already understands the vocabulary of that measurement, then the task is not to beg a chatbot to sound like you in a sentence or two of pleading. The task is to take a proper measurement and hand it over.

That is the whole idea behind a Timbrel voiceprint. The builder walks you through thirteen dimensions of how you actually write, from where you sit on Biber's involved-to-informational line to the exact tells you want tracked down and removed. It never asks you to describe your voice in adjectives, which nobody can do honestly about themselves. It shows you real passages and watches which way you lean.

What comes out the far side is a file, plain text, that you paste into the system prompt of whatever model you use. It states your coordinates in language the model already knows. It bans the machine's defaults by name. And it protects your genuine quirks, so the one phrase you honestly overuse survives the cull while the borrowed AI tics do not.

It is, put another way, the Federalist Papers method run backwards. Mosteller and Wallace measured a fingerprint in order to find a person. A voiceprint measures a person in order to project that fingerprint forward, onto a machine that would otherwise have had none of its own.

A confession about this post

It was written with a Timbrel voiceprint. The profile is witty, relaxed and measured; it was told to keep the paragraphs uneven, skip the rule of three, and refuse the word delve on sight. Whether it managed all three is honestly your call to make.

Build your own →