BaitCheck

A pattern detector for manipulative outrage framing in media. This is NOT a fact-checker and NOT a political bias judge - it's a tool that flags common outrage-bait patterns to help users think critically about content.

Setup

# Install dependencies
npm install

# Run development server
npm run dev

# Build for production
npm run build

# Start production server
npm start

Open http://localhost:3000 to use the app.

Deploy to Vercel

Or deploy via CLI:

npm i -g vercel
vercel

How It Works

Scoring Pipeline

BaitCheck uses a rule-based scoring system that analyzes text for five categories of outrage-bait patterns:

Category	Weight	Description
Loaded Language	25%	Emotional, inflammatory words (e.g., "disgusting", "evil", "scum")
Absolutist	15%	Certainty/black-and-white language (e.g., "always", "never", "everyone knows")
Threat/Panic	25%	Fear-mongering framing (e.g., "they're coming for", "under attack", "collapse")
Us-vs-Them	15%	Divisive in-group/out-group language (e.g., "those people", "elites", "real Americans")
Engagement Bait	20%	Clickbait/viral patterns (e.g., "you won't believe", "shocking", "must see")

Score Calculation

1. Pattern Matching: Scans text for dictionary phrases in each category
2. Normalization: Counts are normalized per 1,000 words and capped to prevent extreme scores
3. Category Scores: Each category gets a 0-100 score based on pattern density
4. Weighted Sum: Final score combines category scores with their weights
5. Co-occurrence Bonus: +5-10 points if multiple categories score high simultaneously
6. Final Score: Clamped to 0-100

Score Labels

• Low (0-33): Minimal outrage-bait patterns
• Medium (34-66): Moderate presence of manipulative framing
• High (67-100): Heavy use of outrage-bait techniques

Project Structure

src/
├── app/
│   ├── api/
│   │   └── analyze/
│   │       └── route.ts    # POST /api/analyze endpoint
│   ├── layout.tsx          # App layout
│   ├── page.tsx            # Main UI
│   └── globals.css         # Styles
└── lib/
    ├── extract.ts          # URL fetching & text extraction
    └── score.ts            # Scoring pipeline & dictionaries

API

POST /api/analyze

Analyzes a URL for outrage-bait patterns.

Request:

{
  "url": "https://example.com/article"
}

Response:

{
  "success": true,
  "score": 45,
  "label": "Medium",
  "reasons": [
    "Heavy loaded language and emotional appeals (\"outrageous\", \"disgusting\")",
    "Strong us-vs-them divisive framing (\"they\", \"elites\")"
  ],
  "highlights": [
    { "start": 123, "end": 133, "category": "loadedLanguage", "text": "outrageous" }
  ],
  "signalBreakdown": {
    "loadedLanguage": 65,
    "absolutist": 20,
    "threatPanic": 30,
    "usVsThem": 45,
    "engagementBait": 15
  },
  "title": "Article Title",
  "sourceDomain": "example.com",
  "textPreview": "First 500 characters of extracted text..."
}

Adding New Phrases/Weights

Adding Phrases

Edit src/lib/score.ts and add to the SIGNAL_DICTIONARIES object:

const SIGNAL_DICTIONARIES: Record<SignalCategory, string[]> = {
  loadedLanguage: [
    // Add new phrases here
    "new phrase",
    "another phrase",
    ...existingPhrases
  ],
  // ... other categories
};

Adjusting Weights

Edit the CATEGORY_WEIGHTS object in src/lib/score.ts:

const CATEGORY_WEIGHTS: Record<SignalCategory, number> = {
  loadedLanguage: 25,  // Adjust these values
  absolutist: 15,      // Must sum to 100
  threatPanic: 25,
  usVsThem: 15,
  engagementBait: 20,
};

Adjusting Category Sensitivity

Modify the multipliers in the signalBreakdown calculation:

const signalBreakdown: SignalBreakdown = {
  loadedLanguage: Math.min(100, normalizedCounts.loadedLanguage * 10),  // Adjust multiplier
  // ...
};

Security Features

• SSRF Protection: Blocks private IP ranges and localhost
• Protocol Restriction: Only allows HTTP/HTTPS
• Size Limits: 5MB max response, 15s timeout
• Content Type Validation: Only processes HTML/text content
• In-Memory Caching: Prevents repeated fetches of same URL

Limitations

• Cannot bypass paywalls
• May not extract text from JavaScript-heavy sites
• Pattern matching is keyword-based (not semantic)
• Scores are heuristic, not definitive
• Some legitimate content may score high if discussing outrage topics

License

MIT