How the Humanizer Works
The Humanizer rewrites AI-generated text so it reads as if a person wrote it, preserving your meaning, facts and formatting while changing the register.
What it does
AI-text detectors don't measure who wrote a text, they measure its style: cadence, uniform sentence length, and a vocabulary of telltale clichés. The Humanizer keeps your content fixed and repaints the writing style into a more human register. It never invents facts or changes your argument — Markdown, math, numbers and names are preserved; only the phrasing and rhythm change.
The pipeline, step by step
1. Register transfer
The model rewrites your text once, holding meaning, facts and markdown constant while sharply varying sentence length and openings, preferring plain wording, and dropping AI constructions.
2. Finding the tells
A second model, from a different family than the one that wrote the draft, quotes the exact machine-writing tells it still carries — uniform sentence length, repeated shapes, em dashes, and the "it's not just X, it's Y" pattern.
3. Targeted refinement, checked locally
The draft is rewritten to fix the quoted tells, and deterministic statistics in our own code measure burstiness, varied openings and repeated patterns, keeping whichever version reads more human. A length-and-meaning safety check discards any rewrite that drifts too far.
The research behind it
The approach follows the most effective text humanizers studied as of 2026:
- Detector-guided decoding (Cheng et al., NeurIPS 2025): generate candidate continuations and keep the one a detector scores as most human.
- RL paraphrase policies (AuthorMist 2025; StealthRL 2026): a paraphraser whose reward balances evasion against meaning preservation.
- Style-transfer humanization, MASH (Gu et al., 2026): hold content fixed and swap a learned style coordinate from machine to human.
The shared finding: swapping style with content held constant is enough to flip a detector's verdict. Today's text-only detectors measure register, not provenance — which is exactly the property the Humanizer edits.
Limitations
- By default we do not run an AI detector and never show a pass/fail AI score; we measure surface structure and push it toward human writing.
- Improving register does not guarantee any specific commercial detector (GPTZero, Originality, etc.) will pass a given text, and detectors change over time.
- Refinement rewrites repeatedly; over several rounds meaning can drift, so always re-read the result.