🕵️ Stylometry and the Pseudpocalypse: How Your Writing Identifies You

Stylometry is the statistical analysis of writing style to identify or characterize an author. The core assumption is that everyone has an idiolect — a personal fingerprint in how they write, made up of habits they're not consciously aware of and can't easily suppress: how often they use function words like 'the', 'of' and 'while', their average sentence length, punctuation patterns, favored word pairs, and thousands of other subtle features. Because these habits are largely unconscious, they persist even when a writer is deliberately trying to disguise their identity or is writing about a completely different subject. The field is also called authorship attribution, and it's old — its first famous success was a 1964 study by Mosteller and Wallace that settled the disputed authorship of the Federalist Papers by analyzing tiny differences in common-word usage, such as the fact that Madison wrote 'whilst' where Hamilton wrote 'while'. What's changed recently is that machine learning, and now large language models, have made it vastly more powerful and accessible.

Given enough text and a defined pool of suspects, yes — often with unsettling accuracy, and the research bears this out. The landmark demonstration is a 2012 study by Arvind Narayanan and colleagues titled 'On the Feasibility of Internet-Scale Author Identification', which showed that stylometry could identify the author of an anonymous piece of text from a pool of 100,000 possible authors with meaningful accuracy — the correct author was the top match around 20 percent of the time, and in the top few matches far more often. That's at internet scale, from a huge suspect pool. With smaller pools the accuracy climbs steeply: studies differentiating around 100 authors have reported accuracy as high as 94 percent. And crucially, the identification survives context-switching — earlier work by Rao and Rohatgi showed that anonymous posts could be linked to a person's known writing even when the two samples were about entirely different topics, and the threat persists across domains, such as linking blog posts to emails. Your writing style is, in a real sense, a biometric.

It's a term popularized by the writer Dynomight for the idea that online pseudonymity is dying — the conjecture that if you put any significant amount of text on the internet under different names, those identities can eventually be linked using only the text. The stronger version is a 'generalized pseudpocalypse': the notion that as identification technology improves, interacting with the world through almost any rich channel will end up revealing who you are. Mask your face and use a voice changer, and you can still be identified by gait or body shape; ban license plates, and cars can be tracked by unique engine sounds or tire wear; lock down your browser and pay only in Monero through three VPNs, and you might still be identified by how you move the mouse or scroll. The writing-style version is the most immediately real, because the tools already exist and are improving fast. Notably, the author observed that a current large language model, given the first thousand words of a draft, correctly identified him — the kind of near-effortless attribution that used to require a specialist and a research paper.

They've collapsed the cost and expertise required, which is the whole reason this shifted from an academic curiosity to a practical privacy threat. Traditional stylometry required building feature sets, training classifiers, and assembling a labeled corpus of candidate authors — real work for a specialist. Large language models can now often guess authorship, or infer an author's traits, from a short sample with no special setup at all, simply by being asked. This 'democratization' is the dangerous part: it takes a previously labor-intensive, expert task and turns it into something automated, streamlined and widely accessible. A capability that a well-resourced adversary once had is now available to almost anyone. And the models keep improving at identifying increasingly obscure writers from increasingly small amounts of text, including across different registers and subjects. The trajectory points toward a world where any substantial body of pseudonymous writing can be linked to its author with minimal effort.

The two most famous cases bookend the technology's history. The Unabomber, Ted Kaczynski, was a domestic terrorist whose identity remained unknown through a long bombing campaign — until his manifesto was published and his brother recognized the writing style and phrasing, prompting the investigation that caught him. That was human recognition of idiolect, before automation. The landmark computational case is J.K. Rowling: in 2013 she published the detective novel 'The Cuckoo's Calling' under the pen name Robert Galbraith, and after a tip-off, computational linguists including Patrick Juola compared the book's common-word and character-n-gram patterns against Rowling's known work and found a strong match. It was enough for the Sunday Times to confront her agent, and she confirmed it. It's worth stressing what that case did and didn't prove — the analysis established that the book was either by Rowling or by someone who wrote very much like her, not a certainty. But combined with the tip, it was decisive. Both cases illustrate the same lesson: sustained writing under a pseudonym is a persistent link to your identity.

No, and this is the most important caveat, one that even enthusiasts insist on. Stylometry is probabilistic, not conclusive. As the computational linguist who worked the Rowling case and others have emphasized, it's far less reliable than something like DNA — a stylometric match tells you the text was likely written by a particular person or by someone with a very similar style, not that it definitively was. Writing style varies within a single author across genres, moods, and years, which is exactly why identification is hard and why results come with error rates rather than certainties. Historically this unreliability, and some high-profile forensic failures, limited stylometry's admissibility as courtroom evidence, though rigorous analyses have gained more acceptance over time. The practical upshot: stylometry is a powerful tool for generating leads and raising or lowering probabilities, and it can absolutely be enough to break anonymity when combined with other clues — but treat any single stylometric 'identification' as strong suspicion, not proof.

Partly, with real effort, and the honest answer is that it's hard and imperfect. There's a research field called adversarial stylometry devoted to exactly this, and it has found some workable defenses. The main approaches are: obfuscation (consciously changing your writing to push its statistical features toward the average, making them less distinctive), imitation (deliberately writing in the style of someone else), and translation (round-tripping text through machine translation to launder stylistic markers, though this degrades quality). Studies have shown that even untrained people, given basic guidance, can reduce an attacker's accuracy substantially — in some cases down toward random guessing — but doing it consistently across a large body of text is genuinely difficult, because your habits reassert themselves the moment you stop concentrating. There are also tools designed to help, historically things like Anonymouth, which flag the features making your writing identifiable. None of this is foolproof against a determined, well-resourced adversary.

Accept that writing style is an attack surface, and design around it rather than assuming a pseudonym protects you. The most reliable defenses are behavioral, not technical. Minimize the volume of text you publish under any identity you want to keep separate — the less you write, the less signal there is to fingerprint, because these methods need a reasonably large sample to work. Keep pseudonymous writing genuinely separate from your real-name writing, and be aware that the link survives even across different topics, so writing about unrelated subjects is not protection. For genuinely high-stakes anonymity — whistleblowing, dissent, sensitive reporting — consider using an intermediary who rewrites your words, writing collectively so no single style dominates, having your text heavily edited, or simply accepting that a sophisticated adversary may be able to identify you and planning accordingly. And combine this with the rest of a real operational-security posture, because stylometry is only one of many channels through which anonymity leaks. The blunt truth of the pseudpocalypse argument is that perfect textual anonymity, at scale, may no longer be achievable — so the goal is raising the cost of attribution, not guaranteeing it's impossible.