Can your writing style identify you?

Yes. Given a few hundred words and a set of candidates, stylometry, the statistical study of writing style, picks the author out far more often than chance, and it does it from the parts of writing you never think about: how often you write "of" and "the", where your commas go, how long your sentences run. It does not need your name, your subject or your opinions. Paste a text into TELLS to see which habits give yours away, in order of how unusual each is; it flattens the mechanical ones with one press and tells you plainly that the rest are yours to change. Nothing is uploaded.

What it matches on

Not the words you would expect. Topic words (names, places, technical terms) change with the subject, so a method that used them would only find people who write about the same things. The measures that survive a change of subject are the ones nobody chooses on purpose:

• Function words. The rate at which you use "the", "of", "and", "that", "which", "very", "quite", "just". A hundred of these, counted per thousand words, is the classic fingerprint, and it has been since the 1960s, when it was used to settle who wrote the disputed Federalist Papers.
• Punctuation. Commas per sentence, semicolons at all, a dash and which dash, brackets, the ellipsis, an exclamation mark, quotation marks straight or curly, one space after a full stop or two.
• Sentence rhythm. The average sentence length and, more telling, how much it varies. Some people write in a steady twenty words; some alternate long and short.
• Small conventions. A spelling convention (colour or color, -ise or -ize), contractions or none, a capital I or a lower one in a hurry, an emoji, a favourite opener ("So,"), a pet phrase.
• Character patterns. Runs of three or four letters, counted blind, which capture spelling, suffixes and typing habits without the method needing to know a language at all.

None of these is a fingerprint on its own. Together, across a few hundred words, they are.

How little it needs

Research on authorship attribution has, on ordinary prose, separated tens of authors from around five hundred words each and done better than chance with far less. The number to hold on to is that a long email, a forum post or a review is enough. A tweet is not, and a method that claims a match from one is guessing. Accuracy falls when the candidates write about different things, when the texts are of different kinds (a letter against a report), and when the writer was deliberately changing style, which is the case that matters here.

The cases people know

In 2013 a crime novel by an unknown first-time author was shown, by comparing its style with the known work of a handful of candidates, to have been written by J. K. Rowling; the analysis matched on word lengths, common words and character patterns, and the publisher confirmed it. The Unabomber's manifesto was recognised by his brother from its phrasing, and linguistic comparison with his letters was part of the case for the search warrant. Newsrooms have used the same methods to test whether an anonymous letter came from the person it seemed to, and courts have heard it as evidence with varying degrees of welcome. It has also gone wrong, and a comparison that looks decisive can be comparing subject rather than author, which is why the number a tool gives you is a place to start asking, not an answer.

What can be flattened, and what cannot

Some tells are mechanical: they live in the characters, not in the thought, and can be normalised without touching a sentence. TELLS flattens these with one press: curly quotes to straight, every dash to one kind, double spaces to single, the three-dot ellipsis to one form, a lower-case "i" to a capital, emoji removed, one spelling convention chosen and applied, contractions expanded. That takes away the cheap giveaways, the ones a casual reader or a quick script would use.

The rest cannot be flattened, and anything that says it can is selling something. Which words you choose, how your sentences move, what you put first, how you argue: these survive every mechanical change, and the research on adversarial stylometry is consistent that a determined writer can change them by hand and a tool cannot do it for them. A tool that rewrote your sentences would give the text a style of its own, and a text with a tool's style is as traceable as one with yours: to the tool. TELLS lists the deep habits so you know where to look, and leaves the sentences to you.

Comparing two texts

Put two texts side by side and TELLS builds a profile of each (a hundred function-word rates, a dozen punctuation rates and the sentence rhythm) and gives the cosine similarity between them. Above about ninety-three usually reads as the same hand; below seventy as different; between is between. The score is weak on short texts and on texts about different things, and it is one comparison, not an attribution among candidates. Use it to ask whether a text you are about to post reads like the rest of what you have posted, or whether the two letters on your desk plausibly share an author. Do not use it to accuse anybody.

What actually helps

• Flatten the mechanical tells, because they are free to fix and are what a quick match uses.
• Read the list of deep habits and change the ones at the top by hand: the phrase you lean on, the way you open a paragraph, the sentence length you fall into.
• Do not post the same text, or the same paragraph, under two names.
• Remember that style is one signal. An account's posting times, its subjects and its metadata identify a person at least as well; SCRUB and CLOAK deal with the metadata of the file and the photo, which TELLS does not.