GitHubHelp

WaywardHayward/ai-detector

Edge/Chrome extension to detect AI-generated content using linguistic heuristics

★ 1Forks 0TypeScriptGitHub ↗Compare

README

AI Content Detector

An Edge/Chrome extension that analyzes web page content to detect potential AI-generated text using linguistic analysis.

🎯 How It Works

The extension uses multiple linguistic signals to estimate the probability that text was AI-generated:

Detection Signals (AI Indicators)

Signal What it detects
Em-dash overuse AI (especially Claude) overuses em-dashes (—)
AI phrases 90+ GPT-isms: "delve into", "it's important to note", "leverage", etc.
Formal starters "However,", "Additionally,", "Furthermore," at sentence starts
Robotic patterns "both X and Y", "not only... but also", passive hedging
Low vocabulary diversity Repetitive word choices (low type-token ratio)
Uniform sentences Similar sentence lengths (low "burstiness")

Human Signals (Authenticity Indicators)

Signal What it detects
Casual markers "gonna", "tbh", "lol", informal language
Self-corrections "I mean", "wait", "actually"
High burstiness Varied sentence lengths (natural human rhythm)
Incomplete thoughts Ellipses, trailing off...

📦 Installation

From Source

# Clone
git clone https://github.com/WaywardHayward/ai-detector.git
cd ai-detector

# Install & build
npm install
npm run build

# Load in browser
# 1. Go to edge://extensions/ or chrome://extensions/
# 2. Enable "Developer mode"
# 3. Click "Load unpacked"
# 4. Select the `dist` folder

🚀 Usage

  1. Navigate to any web page with text content
  2. Click the extension icon
  3. Click "Analyze Page" to scan the main content
  4. Or select specific text first to analyze just that selection

📊 Understanding Results

  • Score 0-34: Likely Human Written ✅
  • Score 35-64: Possibly AI Generated 🤔
  • Score 65-100: Likely AI Generated 🤖

Key Metrics

  • Em-dashes: Count per 1000 words (AI: >2, Human: <1)
  • AI Phrases: Common GPT-isms per 1000 words
  • Vocab Diversity: Unique/total words (Human: >50%, AI: <45%)
  • Burstiness: Sentence length variance (Human: >0.5, AI: <0.3)
  • Human Markers: Casual/informal elements found

⚠️ Limitations

This tool uses heuristics and is not definitive:

  • Well-edited AI text may appear human
  • Formal/technical human writing may appear AI-like
  • Short text samples (<200 chars) are unreliable
  • Cannot analyze images, PDFs, or dynamically loaded content

Use as one signal among many, not as proof.

🔧 Development

npm install          # Install dependencies
npm run build        # Build to dist/
npm run typecheck    # TypeScript check
npm run watch        # Watch mode (TODO)

📝 TODO

  • Browser action badge showing quick score
  • Highlight detected phrases in page
  • Options page for sensitivity tuning
  • Export analysis reports
  • ML-based classifier (train on real samples)

📄 License

MIT

Issues