Recently, I came across an interesting idea in the comments and thought I’d share it: you can use a language model to rephrase posts and messages, making them less recognizably yours. It’s not really “encrypting stylometry”—that sounds impressive, but isn’t quite accurate. It’s more about changing some of the writing habits that could make different texts seem like they came from the same person.
The basic idea is simple: take a draft, run it through a local model with clear instructions, and review the result. But keep the limits in mind: rephrasing may 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 runs locally through Ollama and supports many languages and dialects. For English-language text, it should be a more practical starting point than a very small model that might produce awkward rewrites or change the meaning. Ollama lists its size at around 2.5 GB, and offers smaller variants if your computer has limited resources.
You can also try Qwen2.5:3b. It’s an older, compact option that also supports English. TinyLlama is worth experimenting with, but I’d compare its output with a somewhat larger model using the same text.
Keep in mind that “local model” means your request is processed on your computer only if the app is actually using the downloaded model locally. Don’t connect cloud APIs or plugins to the workflow if you don’t want the original text sent to an external service.
And also:
- Don’t try to sound like someone else. Asking the model to imitate a specific dialect or group can produce a caricature. It’s safer to ask it to vary the structure and wording without pretending to be a particular person.
- Don’t publish the rewrite without reading it. Read it aloud and remove phrases that sound like generic AI boilerplate or nothing like your usual voice.
Here’s a starting prompt:
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. And ordinary editing doesn’t automatically make a text unrecognizable to analysis.
So 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.
The basic idea is simple: take a draft, run it through a local model with clear instructions, and review the result. But keep the limits in mind: rephrasing may 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 runs locally through Ollama and supports many languages and dialects. For English-language text, it should be a more practical starting point than a very small model that might produce awkward rewrites or change the meaning. Ollama lists its size at around 2.5 GB, and offers smaller variants if your computer has limited resources.
You can also try Qwen2.5:3b. It’s an older, compact option that also supports English. TinyLlama is worth experimenting with, but I’d compare its output with a somewhat larger model using the same text.
Keep in mind that “local model” means your request is processed on your computer only if the app is actually using the downloaded model locally. Don’t connect cloud APIs or plugins to the workflow if you don’t want the original text sent to an external service.
And also:
- Don’t try to sound like someone else. Asking the model to imitate a specific dialect or group can produce a caricature. It’s safer to ask it to vary the structure and wording without pretending to be a particular person.
- Don’t publish the rewrite without reading it. Read it aloud and remove phrases that sound like generic AI boilerplate or nothing like your usual voice.
Here’s a starting prompt:
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. And ordinary editing doesn’t automatically make a text unrecognizable to analysis.
So 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.