Recently, I came across a user’s comment about this method and wanted to share how it could be put into practice.
The idea is to use a language model to rephrase posts and messages, making the author’s style less recognizable. It’s not really “stylometric encryption”—that sounds impressive, but isn’t quite accurate. It’s more about changing some of the writing habits that might make different texts seem as though they came from the same person.
The basic process is simple: prepare a draft, run it through a local model with clear instructions, and review the result. Keep the limitations in mind: rephrasing can change some stylistic cues, but it doesn’t guarantee anonymity.
I’d start with Qwen3:4b rather than TinyLlama. It’s a compact model that can run locally through Ollama and supports many languages and dialects. For English-language text, it may be a more practical choice than a very small model, which can produce awkward wording or change the meaning. Ollama lists its size at around 2.5 GB, and smaller variants are available for computers with limited resources.
You could also try Qwen2.5:3b, an older compact model that supports English. TinyLlama is worth experimenting with too, but it’s useful to compare its output with that of a larger model using the same text.
“Local model” means your request is processed on your computer only if the app is actually using the downloaded model locally. If you don’t want the original text sent to an external service, don’t connect cloud APIs or plugins to the workflow.
Don’t ask the model to imitate a specific person. Trying to imitate a particular dialect or social group can produce a caricature. It’s safer to ask it to change the structure and wording without pretending to speak like someone else.
And don’t publish the rewrite without reading it first. Read it aloud and remove phrases that sound like generic AI boilerplate or unlike your usual voice.
Here’s a prompt to get started:
Rewrite the text below in natural English.
Preserve the facts, meaning, point of view, and level of certainty.
Do not add new details or make the text more emphatic.
Change the sentence structure, rhythm, and wording. Avoid repetitive
phrases, clichés, and an overly polished or promotional tone.
Do not imitate a specific person, regional dialect, or social group.
Return only the revised text.
If the result still feels too close to the original, you can ask for a second version with a different structure, then choose the best phrasing yourself. I wouldn’t run the same text through the model over and over: after several passes, the meaning can start to drift and the writing may acquire a conspicuous “AI polish.”
Even a well-rephrased text can reveal its author through more than writing style. Shared topics, personal details, recurring expressions, posting times, and links between accounts can all help connect posts. Ordinary editing doesn’t automatically make a text unrecognizable to analysis.
It’s better to think of an LLM as an editor that can reduce some stylistic similarities, not as an anonymization tool. Avoid including unnecessary personal details, make sure the processing is actually local, and check the meaning yourself. Rephrasing is one small OpSec measure—not an invisibility cloak.
Написать Join — чтобы участвовать в обсуждении.
we recently wrote an article about this
we use Ai all the time for this exact reason
anyway another good article
Whats your next article going to be about? any ideas?
Ответы · 1
ZCash and XMR - very interesting cryptocurrencies
Hows the mixing service going?
Ответы · 1
Everything’s great! How’s business going for you?
Ответы · 1
not too bad
what are the rates for your mixer
what other forums do you post on?
Ответы · 1
Also Pitch, Tenebris and Dread :)
We have communities on Tenebris - Justmix and CryptoPrivacy
Ответы · 1
can't stand dread, but I didn't know about Tenebris - you got a link to share?