GitHubContact Support
DocsAPI ReferenceGuidesExamples

Search Documentation

Search through documentation, guides, and resources.

How It Works

Understanding Truuly's AI text humanization technology

Last Updated: January 27, 2025

Overview

Truuly uses advanced natural language processing and GPT-4o to transform AI-generated text into natural, human-like content. Our proprietary algorithm analyzes your text, identifies AI patterns, and reconstructs it to sound genuinely human-written.

The Humanization Pipeline

1

Text Analysis

AI probability detection and pattern recognition

2

Strategy Generation

Custom humanization strategy based on content type

3

GPT-4o Processing

Advanced language model transformation

4

Post-Processing

Final polish and consistency checks

5

Validation

Quality assurance and detection testing

1. Text Analysis Phase

AI Probability Detection

Our system first analyzes your text to determine its AI probability score (5-95%). This involves:

  • Burstiness Analysis: Measures sentence length variation (AI tends to use uniform lengths)
  • Perplexity Scoring: Evaluates vocabulary diversity and word choice patterns
  • AI Pattern Detection: Identifies 150+ common AI phrases and 102 AI-specific words
  • Structural Analysis: Examines paragraph structure, transition usage, and formatting

Common AI Telltale Signs

Words:

delve, crucial, leverage, paradigm, comprehensive

Phrases:

"in today's world", "it is important to note"

Content Type Classification

The system automatically detects your content type to apply appropriate humanization strategies:

AcademicTechnicalCreativeBusinessCasualLegalMarketingGeneral

2. Strategy Generation

Based on the analysis, Truuly generates a custom humanization strategy with 8 key components:

Lexical & Morphological Variation

Replaces AI-preferred words with natural alternatives. Varies word forms and introduces authentic vocabulary diversity.

Sentence Structure Chaos

Introduces burstiness by varying sentence lengths dramatically. Mixes short, punchy sentences with longer, flowing ones.

AI Signature Removal

Eliminates all 150+ AI telltale phrases and 102 common AI words. Removes robotic transition patterns.

Personal Voice Injection

Adds subtle personality markers, contractions, and natural speech patterns. Context-dependent based on content type.

Punctuation Play

Varies punctuation patterns naturally. Uses em dashes, semicolons, and parenthetical asides like humans do.

Semantic Preservation

Guarantees meaning integrity while transforming expression. Critical facts and data remain unchanged.

Context Adaptation

Adjusts strategy based on content type. Academic maintains formality, creative allows experimentation.

Mode-Specific Intensity

Calibrates transformation level: Light (10-30%), Balanced (30-60%), Aggressive (60-80%), Stealth (80-95%).

3. GPT-4o Processing

The generated strategy becomes a sophisticated prompt for GPT-4o. Key features:

Model Selection

  • GPT-4o: Used for high AI probability (>70%) or complex text requiring maximum quality
  • GPT-4o-mini: Used for simple rewrites and lower AI probability content

Temperature Calibration

Dynamic temperature (0.45-0.95) based on mode and content type:

  • • Stealth Mode: 0.85-0.95 (high randomness for maximum variation)
  • • Academic Content: Lower temperature for precision
  • • Creative Content: Higher temperature for variation

Prompt Engineering

Our 1,148-line humanizer engine generates adaptive prompts that:

  • Incorporate all 8 strategy components
  • Adapt to content complexity
  • Enforce specific linguistic patterns
  • Maintain semantic integrity

4. Post-Processing

After GPT-4o transformation, the text undergoes final cleanup:

  • Regex-based removal of 50+ leftover AI patterns
  • Multiple space and punctuation fixes
  • Consistency checks across the document
  • Final formatting normalization

5. Multi-Detector Validation

Truuly validates results using 4 detection engines:

GPTZero

Industry-leading detector with sentence-level analysis

ZeroGPT

High-confidence AI content detection

Originality.ai

Professional-grade AI detection service

Local Heuristic

Proprietary pattern-based detector (no API required)

The system calculates an average AI score across all detectors and provides a confidence rating:

  • Human (0-30%): Very low AI probability
  • Mixed (30-65%): Moderate AI signals detected
  • AI (65-100%): High AI probability

Why It Works

Scientific Approach

Truuly's effectiveness comes from understanding how AI detectors work. Most detectors look for:

  • ✓ Uniform sentence structures (we add variance)
  • ✓ Low vocabulary diversity (we increase perplexity)
  • ✓ AI-specific phrases (we eliminate all 150+)
  • ✓ Predictable word patterns (we introduce randomness)
  • ✓ Formal, robotic tone (we add personality)

Continuous Improvement

Our system is constantly updated with:

  • New AI pattern databases
  • Improved detection algorithms
  • Enhanced humanization strategies
  • Latest GPT model capabilities

Learn More

For Users

For Developers