Stylometry.
How writing style gives authors away.
Every writer has habits they can’t see: how long their sentences run and how wildly that varies, whether they reach for “don’t” or “do not”, which little words carry their thinking, how they punctuate a pause. Stylometry is the discipline, part statistics and part natural language processing, that turns those habits into numbers and the numbers into a fingerprint. It has settled centuries-old authorship disputes, helped catch a domestic terrorist, and unmasked a bestselling novelist. Its newest subject is text nobody wrote at all: the output of AI language models.
We should say up front why we care: our product runs stylometry in production, measuring users’ writing so AI drafts can be held to it. That makes us practitioners, not neutral historians. Everything below is checkable, and the measurements are ones you can reproduce on a free tool that runs in your browser.
Three cases that built the field
The Federalist Papers (1963)
Twelve essays were claimed by both Hamilton and Madison for over 150 years. Statisticians Mosteller and Wallace settled it by counting function words: the small, unconscious words like 'while', 'whilst', 'upon', and 'by'. Hamilton and Madison used them at reliably different rates, and the disputed essays matched Madison. The landmark insight still holds: the strongest authorship signals are the words nobody chooses deliberately.
The Unabomber (1996)
Ted Kaczynski's brother recognized the manifesto's ideas and phrasing, including the idiosyncratic 'you can't eat your cake and have it too'. FBI forensic linguists compared the manifesto's style against Kaczynski's letters, and that linguistic analysis supported the search warrant. A person's fixed expressions and syntactic habits followed him from private correspondence into an anonymous public document.
Robert Galbraith (2013)
When a debut crime novelist was rumored to be J.K. Rowling, researchers Patrick Juola and Peter Millican ran computational stylometry comparing 'The Cuckoo's Calling' against her known work: word-length distributions, common-word frequencies, character n-grams. The tests pointed to Rowling, and she confirmed it. Modern stylometry had unmasked one of the world's most-read authors from surface habits she didn't know she had.
What stylometry actually measures
The features that identify a writer are rarely the memorable ones. Vocabulary can be imitated and topics change, so classical NLP work leans on signals that are frequent, unconscious, and stable: function-word rates, sentence-length distribution and its variance (often called burstiness), punctuation densities, contraction habits, character and word n-grams. Distance measures like Burrows’ Delta then compare a document’s profile against candidate authors’ baselines. Modern approaches add machine learning classifiers on top, but the deep principle is unchanged since Mosteller and Wallace: count what the author doesn’t know they’re doing.
The same principle powers our style-match score: sentence rhythm, contraction rate, punctuation habits, casing quirks, person preference, and question rate, computed from a user’s real writing and compared against every AI draft. It’s stylometry pointed at a friendly problem: instead of asking “who wrote this?”, it asks “does this sound like you?”
The new frontier: fingerprinting the machines
AI-generated text turns stylometry’s question around. Language models are, in a sense, the most prolific authors in history, and they have habits too. When we ran deterministic stylometric analysis over 8,000 passages of the HC3 corpus, the machine’s fingerprint was plain: AI answers were 6.3 times as likely to contain stock phrases, used “additionally” 5.6 times more often, and varied their sentence lengths about 22% less than humans. We also found the fingerprint moves: 2023-era ChatGPT barely used em-dashes, while later models became famous for them. Model style is generational, which means honest measurement has to be redone as models change.
That instability is exactly why stylometry beats black-box AI detection for everyday use. A classifier’s probability score ages badly and can’t explain itself. A measurement (“four stock phrases, uniform rhythm, no contractions”) stays true, stays checkable, and tells a writer what to fix.
Frequently asked questions
What is stylometry in simple terms?
Stylometry is the measurement of writing style: how long your sentences run and how much they vary, which small function words you lean on, how you punctuate, which phrases you repeat. Individually these habits look like nothing. Together they form a fingerprint that statistical and NLP methods can match against known writing.
Is stylometry the same as AI detection?
No. AI detectors are machine learning classifiers that output a probability guess about whether a machine wrote something, and they're wrong often enough to be dangerous. Stylometry measures concrete, checkable features. It can say 'this text has flatter sentence rhythm and more stock phrases than this writer's baseline', which is evidence you can inspect, rather than a verdict you must trust.
Can stylometry identify an anonymous author?
Sometimes, given enough text and a good candidate set. It attributed the disputed Federalist Papers to Madison and unmasked Robert Galbraith as J.K. Rowling. But it's probabilistic: it narrows candidates rather than proving identity, it needs substantial samples, and deliberate imitation or heavy editing can defeat it. Serious practitioners state those limits up front.
How does AI-generated text change stylometry?
It gives the field a new species to study. Large language models have their own measurable habits: our analysis of 8,000 passages found AI text 6.3 times as likely to contain stock phrases and about 22% flatter in sentence-length variation than human text. The same tools that fingerprint humans can fingerprint models, and those fingerprints shift with every model generation.
Run stylometry on any text, right now
The checker highlights measurable tells in your browser, free and unlimited, and nothing is uploaded. Your own writing has a fingerprint. Go look at it.
Open the AI-tell checker