Stylometry: How Your Writing Style Can Identify You

Additional resource for the lesson “Writing Style as a Fingerprint (Stylometry)” — Online Anonymity course

Stylometry is the study of writing style as something you can measure: the words you reach for, the way you punctuate, how long your sentences run. Together these habits can act like a fingerprint for text. This resource explains how it works, why language models make it cheaper to apply, and what it means if you write anything under a pseudonym.

A. Why this matters

In this course, privacy is deciding what others know about you; anonymity is acting without what you do being linked to who you are. Earlier resources looked at links your devices and browsers create: an IP address, a cookie, a browser fingerprint. This one looks at a link you create yourself, every time you write.

Writing under a pseudonym is common and legitimate: a review of a local restaurant, a question on a forum, a blog about a hobby. A pseudonym separates a text from your name. It does not separate the text from the way you write, and the way you write can be compared with other texts that do carry your name.

This does not mean every pseudonymous post can be traced to its author; style is one more link: usually weak on its own, stronger when it is combined with what a text says about your life.

What the sources say

The Council of Europe’s 2003 Declaration on freedom of communication on the Internet links anonymity to free expression: member states “should respect the will of users of the Internet not to disclose their identity.” It adds that this does not prevent states from tracing those responsible for criminal acts, in accordance with national law. The Electronic Frontier Foundation (EFF), a digital rights organisation, puts the everyday version simply: “Many people don’t want the things they say online to be connected with their offline identities.”

Measuring style to identify an author is not new. In a study published in 1963, the statisticians Frederick Mosteller and David Wallace examined twelve Federalist papers whose authorship was disputed between Alexander Hamilton and James Madison. Some of their best evidence came from ordinary words: Hamilton used “upon” about three times per thousand words, Madison roughly once in every six thousand. They concluded that Madison, rather than Hamilton, wrote all twelve.

In 2013 the computational linguist Patrick Juola, asked by The Sunday Times, compared a crime novel published under the name Robert Galbraith with books by four candidate authors. He looked at word lengths, the 100 most common words, short sequences of letters and pairs of adjacent words. J.K. Rowling came out as the most likely candidate, and Juola was careful to say that the analysis pointed to her rather than proved it.

Publishing under a pen name is a choice the law recognises: in Italy, for instance, copyright law (article 21 of Law 633/1941) gives the author of a pseudonymous work the right to reveal themselves at any time, so the decision stays theirs, and in 2021 the European Court of Human Rights (Standard Verlagsgesellschaft mbH v. Austria (no. 3)) recognised that online anonymity helps the free flow of opinions and ideas, though it is not an absolute right. Stylometry matters because it can take that decision out of the author’s hands.

The scale has grown since then. A 2012 study by Arvind Narayanan and colleagues, presented at the IEEE Symposium on Security and Privacy, tested author identification across up to 100,000 bloggers: “In over 20% of cases”, the classifiers named the right one. In 2024, researchers at ETH Zurich showed that large language models can infer personal details such as “location, income, sex” from what people write. The same year, a peer-reviewed study published in the Findings of the EMNLP 2024 conference tested language models on telling whether two texts share an author, and on picking the author among 10 or 20 candidates. It reports that they managed both “without the need for domain-specific fine-tuning”, that is, without being specially trained for the task.

B. Key concepts: how stylometry works

Five ideas, from the habits themselves to what machines can now do with them.

Style is a set of countable habits

Writing style, seen this way, is data: how often you use certain words, how long your words and sentences are, which punctuation marks you favour, how you open and close a message. Researchers who built a software package for this work describe their field as “the quantitative study of writing style”. The analysis compares these counts across texts and asks which known author a new text most resembles.

Why it matters to you: Most of these habits are not conscious choices, which is why they stay the same when you change the name at the top.

The small words carry the most

It sounds backwards, but the most revealing words are often the least interesting ones: “the”, “of”, “upon”, “while”, “however”. Mosteller and Wallace called them function words: articles, prepositions and conjunctions whose rates stay fairly stable whatever the topic, unlike words such as “war”. The research package mentioned above takes the most frequent words as its default starting point.

Why it matters to you: A change of subject changes these habits very little: a book review and a work email can share them.

A writeprint is the combination, not one quirk

Here, “writeprint” simply describes the whole pattern, by analogy with a fingerprint. The researchers Ahmed Abbasi and Hsinchun Chen used the name Writeprints in 2008 for a technique “for identification and similarity detection of anonymous identities”. One habit, such as a fondness for semicolons, is shared by many writers. Many habits together, from word lengths to favourite connecting words to the way you use brackets, form a profile that is much rarer. The 2012 study combined more than a thousand such features, from punctuation to function words.

Why it matters to you: As with a browser fingerprint, it is the bundle, not any single habit, that narrows the crowd.

Authorship attribution is a probability, not a proof

Authorship attribution means estimating who, among a set of candidates, most likely wrote a text. It works best when the list of candidates is short and there is plenty of text from each. Juola wrote that the most likely candidate was “probably Rowling”, and acknowledged that someone could write in a similar style by accident or on purpose. In the 2012 study, letting the system decline to guess when it was unsure raised the share of correct top guesses from 20% to over 80%, at the cost of answering less often.

Why it matters to you: A style match is a clue, not an identification. It matters most where someone already has a short list of possible authors, such as colleagues, classmates or members of a small community.

Language models make the analysis cheaper

Comparing styles used to need specialist software and some expertise. Large language models, the systems behind AI chatbots, can read ordinary text and pull out clues on their own. The ETH Zurich study found that they did so at a fraction of the cost (100 times) and time (240 times) that people needed. The EMNLP study adds that they can explain their answer through “a detailed analysis of linguistic features”, the kind of habits described above.

Why it matters to you: Your habits have not changed; the effort needed to compare them has, so assuming nobody would bother is less safe than it was.

C. A practical example: Sara’s book reviews

Sara is 28 and teaches at a secondary school. During the week she writes a newsletter for her class’s parents and plenty of emails to colleagues. In her own time she posts book reviews on a reading forum under a nickname. She has nothing to hide; she simply likes keeping her reading life apart from her working one.

One afternoon a colleague mentions a review she enjoyed on that forum and adds, half joking, that it sounded just like one of Sara’s newsletters. Sara is not alarmed, but she is curious. She had never mentioned the school, her town or her name. What had her colleague noticed?

That evening she put a newsletter and two reviews side by side, and the overlap was easy to see. She opened with “So,” in all three. She used dashes for asides where others use brackets, ended paragraphs with a short question, and wrote “a little” where others write “a bit”. Her favourite phrase, “for what it’s worth”, appeared twice. One review also mentioned her rainy walk to work past the school gates.

What changed when Sara looked at her own style

• She saw that the strongest link was not one word but a combination: the same opening, the same dashes and the same phrase, together.
• She noticed that the line about her walk to work linked the review to her far more directly than any punctuation did, and she edited it out.
• She decided that the rest was fine. A colleague recognising her taste in novels is not a problem for her, and she does not want to write her reviews in a voice that is not her own.

If she ever writes something where the separation matters more, she now knows that content and style both count. Keeping separate identities consistent is a topic the course returns to at a later level.

D. Try it yourself: compare two of your own texts (15 minutes)

Four short steps, with pen and paper or a blank document. Use your own texts only, and keep your notes to yourself: the exercise works without uploading anything to a website or a chatbot.

Step 1 — Pick two texts (3 minutes)

• Choose one text written under your own name, such as an email or a post on a named account, and one written somewhere else, or simply two texts on different subjects.
• Aim for a few paragraphs each: longer texts show more.
• Avoid texts that contain other people’s private information.

Step 2 — Look at words and phrases (4 minutes)

• Underline how each text opens and how it closes.
• Note the small connecting words you reach for, such as “so”, “actually”, “just”, “really” or “however”.
• Circle any phrase or expression that appears in both texts.

Step 3 — Look at punctuation and rhythm (4 minutes)

• Mark dashes, brackets, semicolons, exclamation marks, ellipses and emoji.
• Count, roughly, the words in five sentences of each text. Are your sentences mostly short, mostly long, or a mix?
• Note spelling and formatting choices: “email” or “e-mail”, numbers in words or figures, capital letters at the start of messages or not.

Step 4 — List five habits and decide (4 minutes)

• Write down five habits that appear in both texts and that you think are least common.
• Next to each, write “fine” or “worth knowing”.
• Add any detail of content, such as your job, your town or a routine, that links a text to you more directly than its style.

The aim is awareness, not disguise.

E. Videos, articles and further resources

Institutional sources first, then the research behind the examples in this resource. All in English.

Rights and context

Council of Europe — Declaration on freedom of communication on the Internet (2003)
Principle 7 sets out respect for users who choose not to disclose their identity, alongside the lawful tracing of criminal acts.
https://rm.coe.int/16805dfbd5

Electronic Frontier Foundation — Anonymity
Why anonymous and pseudonymous speech matters, and the EFF’s work to protect it.
https://www.eff.org/issues/anonymity

Research on authorship

CSLI Publications (Stanford University) — Inference and Disputed Authorship: The Federalist
The reissued 1964 book in which Mosteller and Wallace used frequent words to settle the disputed Federalist papers.
https://web.stanford.edu/group/cslipublications/cslipublications/site/1575865521.shtml

Language Log (University of Pennsylvania) — Rowling and “Galbraith”: an authorial analysis
Patrick Juola’s own 2013 account of the four tests he ran, and of their limits.
https://languagelog.ldc.upenn.edu/nll/?p=5315

Princeton University — On the feasibility of internet-scale author identification
The 2012 IEEE Security and Privacy study of author identification across 100,000 bloggers.
https://collaborate.princeton.edu/en/publications/on-the-feasibility-of-internet-scale-author-identification/

The R Journal — Stylometry with R: A Package for Computational Text Analysis
A European research team (Polish Academy of Sciences, Jagiellonian University, University of Antwerp) presents a software package for analysing style.
https://journal.r-project.org/articles/RJ-2016-007

Language models, privacy and authorship

ETH Zurich, SRI Lab — Beyond Memorization: Violating Privacy Via Inference with Large Language Models
The 2024 study showing that language models can infer personal details from ordinary text.
https://eth-sri.github.io/publications/staab2023beyond

ACL Anthology — Can Large Language Models Identify Authorship?
A peer-reviewed 2024 study (Findings of EMNLP 2024) on language models verifying and attributing authorship without task-specific training.
https://aclanthology.org/2024.findings-emnlp.26/

Links checked in November 2026. If an address changes, searching for the title on the organisation’s site usually finds it.

F. The Cyber Welfare Framework: Skills, Awareness, Secure Behaviour

This resource sits on the Awareness pillar at level FL2 – Beginner: noticing that the way you write is a link, before deciding where it matters.

Skills

• Naming the habits that style analysis measures: frequent words, punctuation, sentence length, openings and closings.
• Comparing two of your own texts and listing the habits they share.
• Spotting the content details in a text that link it to you more directly than its style.

Awareness

• Understanding that a pseudonym hides your name but not your writing habits.
• Knowing that authorship attribution gives probabilities, and works best with a short list of candidates.
• Recognising that language models lower the effort needed to compare texts and to infer details from them.

Secure Behaviour

• Re-reading a text you post under a nickname for details about your life you did not mean to include.
• Deciding, in proportion, which of your pseudonymous texts genuinely need to stay separate from your name.
• Keeping other people’s writing out of AI tools when you are curious about who wrote something.

G. Questions to sit with

• Which of your texts would you be comfortable seeing linked to your name, and which would you rather keep apart?
• Who already reads enough of your writing, at work, at school or among friends, to recognise it without a name?
• If an analysis said a text was “probably” yours, how much would that word mean, and to whom?

H. What to do now

The recommendations (R) and security measures (MS) from the Cyber Welfare database that apply most directly here.

No code in the database deals with writing style directly, so the advice below is given without one.

1. Before you post under a nickname

• Re-read for details about your life: your job, your town, your routines, the names of people around you. They usually link a text to you more than its style does.
• Decide which pseudonymous texts really matter. A book review and a personal testimony need different degrees of separation.

Minimum commitment: next time you post under a nickname, re-read it once and remove one detail about your life you did not mean to share.

2. Keeping the limits in view

• Be wary of claims that a tool can “prove” who wrote a text, and of promises that a text can never be linked to you.
• If you are curious about who wrote something, do not paste other people’s private writing into an AI tool to find out.

Minimum commitment: this week, do the two-text comparison in section D once, and keep your list of five habits for yourself.

In short

• Stylometry measures writing habits such as frequent words, punctuation and sentence length, and compares them across texts.
• A pseudonym hides your name, not your writing habits; the combination of habits is what can point to you.
• Authorship attribution gives probabilities, and language models now make it cheaper to apply, often together with the details a text mentions.
• You do not have to change how you write. Knowing your habits, and what your texts say about your life, lets you choose where separation matters.

Related resources in this course

• Reverse Image Search Yourself: Who Can Find You from a Photo
• Protect Your Identity from Deepfakes: Limit, Notice, Act
• How to Use AI Chatbots Anonymously, and Where It Stops
• Re-identification: Why a Few Details Can Identify You

Discover more companion resources from the online courses of the Protect Your Digital Privacy programme.