How HumaraGPT works
From pasting your text to the final detection-proof output — every step explained.
The humanization process
What happens when you click “Humanize” — step by step.
You paste your AI-generated text
You paste or type your text into the editor. You can choose an engine mode (Fast, Standard, or Stealth) depending on how aggressive the rewrite should be. The frontend sends your text, selected engine, and any style preferences to the backend API.
Per-sentence AI risk scoring & pre-detection
Before rewriting, the backend computes per-sentence AI risk scores using signal-based analysis — detecting AI phrases, vocabulary patterns, structural markers, and low perplexity. Each sentence gets an individual risk tier (low, medium, high, critical) that determines how aggressively it will be rewritten. Simultaneously, multiple AI detectors establish baseline “before” scores.
Detectors checked
Adaptive engine selection & domain analysis
Based on the risk profile, the adaptive formula selects the optimal engine from 20+ options — Ghost for stealth, Vine for sequential rewriting, Whisper for detectability, Humarin for anti-GPTZero, and more. Domain-aware processing detects the subject area and loads 50+ protected compound terms to preserve technical accuracy.
Convergence loops & structural rewriting
The selected engine applies adaptive convergence loops — iterating sentence-by-sentence until each one exceeds its minimum change threshold (40-65% depending on risk tier). Natural burstiness is introduced through varying sentence lengths and complexity. A cumulative document change target of 65-85% ensures thorough transformation while preserving meaning above 95%.
OHAS calibration & human baseline replication
The rewritten text is calibrated using OHAS (Overall Human Authenticity Score) — 12 linguistic metrics including sentence variability, tense naturalness, informality ratio, burstiness, and conjunction opener rate. The Human Baseline Replication framework then injects natural hedging, varied sentence openers, and minor errors that match real human writing patterns. Tone is calibrated to match your target style — academic, professional, or conversational.
Post-detection verification & results
The humanized text is scanned again through the same detectors to confirm it now scores as human-written. The API returns both the “before” and “after” detection scores so you can see exactly how much the score improved. The editor highlights per-sentence AI probability so you can fine-tune individual sentences if needed.
What you get back
The AI detection process
How the standalone AI Detector works when you check any text.
Step 1
Submit text
Paste any text into the AI Detector to check it against multiple detectors.
Step 2
Multi-detector scan
7+ detectors run in parallel — Turnitin, GPTZero, Originality.AI, Copyleaks, and more.
Step 3
Aggregated results
Overall AI/human scores, per-detector breakdown, and a clear verdict.
Try it yourself
Paste your text and see the full pipeline in action.