TL;DR
Style-Aware Anonymization is a method that first asks an LLM to build a "style passport" of an author (sentence length, word choice, tone, punctuation) based on a few text samples, then rewrites a new text to strip out exactly those markers while keeping the meaning intact.
People assume that removing names and obvious details is enough to stay anonymous. But author-detection models catch people by how they write: how they build sentences, which words they favor, how long their sentences tend to run. Even "noisy" anonymization methods (based on differential privacy) either leave these style fingerprints intact or turn the text into mush — and still don't guarantee anonymity.
The method solves this in two steps: first, the LLM reads several of your texts and describes your style in words (not numbers, but plain human language: "short, clipped sentences," "loves bureaucratic phrasing"), then it rewrites a new text, deliberately breaking exactly those patterns instead of blindly rewriting everything.
Method Blueprint
STEP 1 (separate request): Style profiling
Input: 3-5 samples of the author's text
LLM analyzes 4 parameters: sentence length, vocabulary, tone, punctuation
→ Output: a text-based style profile (human-readable, not numeric)
STEP 2 (separate request): Rewriting with the profile in mind
Input: new text + style profile from step 1
LLM rewrites the text, specifically neutralizing those style markers
→ Output: anonymized text with the meaning preserved
Example in Action
Task: An employee wants to leave an anonymous review of their employer on Glassdoor or Indeed, but worries coworkers will recognize their writing style — they're known in work chats for constant jokes with signature phrases and long, elaborate sentences.
Prompt (step 1 — profiling):
Here are 5 texts I wrote (chat messages, posts, comments):
{text 1}
{text 2}
{text 3}
{text 4}
{text 5}
Analyze my writing style across four parameters:
1. Sentence length and structure
2. Word choice and vocabulary (favorite words, jargon, stock phrases)
3. Tone (ironic, serious, emotional, etc.)
4. Punctuation (heavy use of ellipses, exclamation points, parentheses?)
Describe each parameter in 2-3 plain-language sentences.
Prompt (step 2 — rewriting):
Here's a profile of my writing style:
{profile from step 1}
Rewrite the text below to remove every trait described in the profile above
(change the sentence length, replace the characteristic vocabulary, shift the tone,
remove the characteristic punctuation). Keep the meaning and all facts fully intact.
Text: {employer review}
Result: In step 1, the model produces a text description of the style — something like "the author writes long, compound sentences, often uses irony, and loves dashes and ellipses." In step 2, the model returns a rewritten review — shorter, drier, stripped of the signature turns of phrase, but with the same facts and complaints. Coworkers reading only the text won't be able to identify the author by writing style, even though the content stays the same.
Why This Works
Models that try to identify the author of an anonymous text ("who wrote this") don't latch onto facts or topics — they latch onto habits: sentence length, favorite words, the rhythm of punctuation. That's a persistent "fingerprint" that doesn't disappear just because you removed the name and obvious details.
Random "noising" of text (the differential-privacy approach) hits everything indiscriminately — either not hard enough, so the style stays recognizable, or too hard, turning the text into a meaningless word salad. It's like trying to disguise someone's voice by just adding static to the recording — either the voice is still recognizable, or the recording turns into pure hiss.
LLMs are good at describing patterns in words and following targeted instructions. The method leverages exactly that: instead of randomly wrecking the text, the model first names the author's specific habits, then takes aim at those and leaves everything else (meaning, facts, structure) untouched.
Levers you can pull: - Number of text samples (the paper used 5) → fewer samples make the profile less precise; more samples improve accuracy but take longer to collect - Which parameters to include in the profile → for short texts (reviews, comments), you can focus on sentence length alone — that's often enough - Explicit profile vs. just "rewrite it neutrally" → an explicit profile works noticeably better than simply asking the LLM to "make the text neutral" without specifying what exactly to change
Prompt Template
STEP 1 — Style profiling:
Here are {number} texts I wrote:
{author's texts}
Analyze my writing style across four parameters:
1. Sentence length and structure
2. Word choice and vocabulary
3. Tone
4. Punctuation
Describe each parameter briefly, in plain language.
---
STEP 2 — Rewriting to neutralize style:
Here's a profile of my writing style:
{profile from step 1}
Rewrite the text below to remove every trait described in the profile
(change sentence length, vocabulary, tone, punctuation).
Keep the meaning and all facts fully intact — don't shorten the content.
Text: {text to anonymize}
🚀 Quick Start — paste this into chat:
Here's a template for anonymizing text by writing style. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which texts to use as style samples and which specific text needs to be anonymized — without that, it can't build an accurate style profile.
Limitations
⚠️ No guarantee of full anonymity: the method cuts author-identification accuracy by 60-70%, but doesn't eliminate it entirely. For short texts (reviews, comments), the risk of de-anonymization is higher than for long ones.
⚠️ Leftover content can still give you away: if the attacker knows the context (say, that the person works at that specific company), the content itself can expose the author even if the style is completely changed.
⚠️ You need samples of your own style: the method requires 3-5 texts you've written previously to build the profile. Without them, the method simply won't work.
⚠️ Short texts are harder to anonymize: for short reviews, the model relies almost entirely on sentence length — if the author consistently writes short, uniform phrases, that's harder to hide than in a long text with varied structure.
How It Was Researched
The team used two datasets: long blog posts (10 authors, 1,500 texts per author) and short Google reviews (the 9 most active reviewers from Illinois). Each author's texts were anonymized using several methods — differential privacy (noise), simple paraphrasing, the competing ALISON method, and their own style-aware approach.
They then tested whether an authorship classifier (a specially trained model) could still identify the author of the anonymized text, and how well the meaning and readability held up (measured via text similarity and "naturalness" as judged by the GPT-2 language model).
The interesting finding: noise-based methods either fail to protect or wreck the text. With light noise, the author is still identifiable; with heavy noise, the text turns into garbage with a "perplexity" (a measure of textual unnaturalness) in the thousands, versus a normal range of 40-100. The style-aware approach cut author identifiability by 60-70% while keeping the text readable (perplexity close to the original) — because the method doesn't randomly break the text, it specifically targets stylistic habits.
A curious detail: for short reviews, a profile based on sentence length alone worked almost as well as the full four-parameter profile — because in short texts, sentence length is the strongest author "fingerprint."
Resources
I Am No One: Style-Aware Paraphrasing for Text Anonymization. Ahmed Sohair Khan, Estrid He, Monica Wachowicz, Elham Naghizade — RMIT University, Melbourne, Australia. Code and materials: github.com/ahmedsohair/SAPTA26