# Voiceprint · Humanizer Extension

Optional add-on to the voiceprint skill. Load it only when someone is judged by
an AI detector and asks for it.

This lives in its own file on purpose. The core skill is about a person's
voice, which is stable for years. Detectors, models and evasion techniques
change every few months. Keeping the churn out here means the core stays
trustworthy and this file can be replaced without touching it.

Version 1, 01.08.2026. If you are reading this much later than that, treat the
numbers as historical and re-measure before quoting them.

---

## Read this before you use any of it

A voice profile reliably moves surface detectors. It does not move GPTZero.

Measured on 30.07.2026 by the author: the same content, written through a
voice profile, went from 100% to 0% on ZeroGPT and from 55% to 0% on QuillBot.
GPTZero did not move at all. Twelve AI-written variants were tested one
variable at a time (5 style variants, 3 cross-model rewrites through Gemma,
Qwen, DeepSeek and Mistral, 1 long version with concrete anchors and
German/English code-switching, 1 English work email, 1 German work email, 1
high-temperature generation). All twelve scored 100% AI. The control, the
author's own untouched German internal email from 2011, in the same register
as the German AI variant, scored 0%.

So the honest summary: surface detectors score style, and style can be moved.
Deep classifiers score something no prompt reached in twelve tries across two
languages, four model families and four registers. Only changing the input,
the person's own sentences, changed the answer.

Never promise a score. Never sell "undetectable". If someone needs a
guaranteed GPTZero pass on fully machine-written text, tell them plainly that
no free prompt method delivers that, and neither do most paid tools.

## How paid humanizers actually work

Five parts: linguistic analysis, a perturbation engine, a fine-tuned rewriting
model, a detector-in-the-loop optimizer, post-cleanup.

The engine is the fine-tuned model, and a prompt cannot reproduce it. Two of
the five parts need no trained model and are free to copy: cross-model
rewriting and the detector-in-the-loop. That is what this file is. Reality
check: GPTZero catches roughly 93% of humanized text, so even the paid tools
mostly fail against it.

## The loop

1. **LENGTH FIRST.** Under about 100 words, detectors swing to all-or-nothing.
   Get to 250+ words of real, specific content before optimizing anything.
2. **CROSS-MODEL LAUNDERING.** Rewrite the draft in a different model family
   than the one that wrote it: Mistral, Qwen, Gemma, DeepSeek, or a local
   model. A different token distribution raises perplexity, which is what
   perplexity detectors score.
3. **DETECTOR-IN-THE-LOOP.** The real mechanism. Paste into the detector, take
   only the flagged sentences, rewrite just those: "these read as AI, rewrite
   only these to break the pattern, keep meaning and facts." Re-score. Up to
   three rounds, then stop. More rounds only drift the meaning.
4. **RECURSIVE PASS.** If it is still high, run the whole thing through a
   second different model once. Diminishing returns after that.
5. **SANITY READ.** Cross-model passes garble a line or shift a fact. Read
   once, fix the one broken line. Never let facts drift.
6. **FOR GPTZERO SPECIFICALLY.** Skip steps 2 to 4 and mix in 30 to 40% of the
   person's own real sentences. Co-writing is the only lever that moved it in
   testing, and it is the core method of the main skill, not a trick.

## Aggressive rewrite prompt

For cross-model laundering and for rewriting flagged sentences:

> Rewrite to break AI-detection patterns while keeping the meaning and every
> fact. Increase unpredictability: vary sentence length violently (some 2 to 3
> words, some long), mix registers, choose uncommon-but-natural words, break
> some parallelism, allow one slightly messy tangent, start mid-thought, do not
> end on a question. Never use [ban-list]. Invent nothing. Output only the
> rewrite.

## What this loop costs you

Every pass moves the text further from the person's voice. That is the trade,
and it should be stated out loud every time: you are exchanging recognition
for a score. After three rounds a text usually passes the surface detectors
and no longer sounds like anyone.

If the person's goal is to sound like themselves, run the core skill and stop.
If their goal is a number on a screen, run this and accept what it costs. Do
not pretend the two are the same thing.

## Never do

Intentional typos (costume, not voice). Synonym spinning (more machine, not
less). Humanizer chains, one tool feeding another (fact drift plus
re-flattening). Invisible Unicode tricks (fragile, zero quality). Cranked
temperature (errors, not personality). Any claim about a guaranteed score.

---

Voiceprint by Engin Senli. https://enginsenli.com/voiceprint
MIT licensed, see LICENSE. Free to use, change and share, including commercially.
Keep this notice. Provided as is, without warranty of any kind. Whatever the AI
writes with this stays your text and your responsibility.