How It Works
Text Humanize is not a synonym spinner. This page explains the exact technical approach we use to make AI-generated text undetectable — and why it works where simpler tools fail.
AI detectors identify machine-generated text by measuring statistical properties that differ between human and AI writing:
Humans write sentences of dramatically different lengths. AI tends to write at a consistent, "medium" length. Detectors measure this variance — a low burstiness score is a red flag.
AI models always choose the statistically most likely next word. Human writers make unexpected word choices. Low perplexity (too predictable) = AI flag.
AI reuses the same vocabulary at a higher rate than humans. Moving Average Type-Token Ratio below ~0.68 is a common AI signal.
AI overuses formal connectors like "Furthermore," "Additionally," and "Moreover." Humans use them far less frequently and more naturally.
AI uses epistemic hedges ("perhaps," "arguably," "seemingly") at a higher rate than humans, creating a tell-tale cautious tone.
Certain words ("delve," "leverage," "paradigm," "undoubtedly") appear at disproportionately high rates in AI output. Their frequency is a direct signal.
Each piece of text goes through three sequential stages before you receive the output.
We run your text through our TextMetrics engine, which calculates all 9 linguistic signals: burstiness, sentence length CV, MATTR, hedging density, Flesch variance, transition word density, AI vocabulary fingerprints, Zipf compliance, and entity coherence. The engine identifies the dominant flaw — the single metric that deviates most from human norms. This flaw becomes the primary target of the rewrite.
We construct a precision prompt that instructs the LLM to fix the dominant flaw — not a generic "make this sound human" instruction. For example, if burstiness is the issue, the prompt explicitly targets sentence length variation. If AI vocabulary is the problem, the prompt focuses on replacing those specific words.
Model selection is adaptive: LLaMA 3.1 8B handles standard requests (fast, cost-efficient). LLaMA 3.1 70B is automatically selected for texts over 1,000 characters, aggressive intensity, or academic purpose — where higher reasoning quality is needed.
Before sending to the model, we extract and protect critical tokens — URLs, citations, proper nouns, numbers, and quoted passages — so they are never rewritten. After the LLM responds, these tokens are restored exactly.
The output is scored on a 0–100 quality scale measuring three factors: length ratio (the output should be 0.8–1.2× the input length), semantic similarity (word overlap and entity retention), and readability improvement.
If the quality score falls below 78, a second corrective pass runs automatically. The corrective prompt is tailored to address whatever specific issue caused the low score. This two-pass system is what produces consistently reliable output rather than hit-or-miss results.
Our TextMetrics engine calculates these signals on both input and output, giving you before/after visibility into what changed.
Burstiness (25 pts)
Sentence word-count variance. Formula: (σ − μ) / (σ + μ). Negative = AI-like uniform rhythm.
Sentence Length CV (22 pts)
Coefficient of variation in sentence lengths. Below 0.15 indicates mechanical consistency.
MATTR (20 pts)
Moving Average Type-Token Ratio in a 50-word sliding window. Human norm: 0.72+.
Hedging Density (12 pts)
Count of epistemic markers per 100 words. AI scores above 0.40 per 100 words.
Flesch Variance (8 pts)
Readability variance across paragraphs. AI maintains unnaturally consistent readability.
Transition Word Density (7 pts)
Frequency of formal connectors. AI uses them at 3–5× the human baseline rate.
AI Vocabulary Fingerprints (6 pts)
50+ characteristic AI words tracked by frequency. 3+ occurrences triggers a flag.
Zipf Compliance (5 pts)
R² fit of word frequency distribution to Zipf's law. Human text deviates — AI text doesn't.
Entity Coherence (4 pts)
Jaccard similarity of named entities between consecutive sentences. AI lacks topical flow.
Every week, we run a standardised test corpus through our humanizer and then through each major detector. The corpus includes text generated by GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro across three content types: academic essays, marketing copy, and general articles.
We record the percentage of documents that score below the detector's AI-flagging threshold. Our published lower AI detection scores reflects the average across all content types and all detectors tested. Some categories perform better (marketing copy: ~99%) and some are harder (short academic excerpts: ~94%).
When a detector updates its algorithm and our bypass rate drops, we update our prompts within 1–2 weeks. This is why we publish "Updated weekly" and not a fixed claim.
Free, no login required. See the before/after metrics on your own text.
Start Humanizing Free