A Chrome Extension for Real-Time Statement Verification

Leverage on-device AI to identify questionable claims, debated facts, and hyperbole as you browse the web — all without sending your data to the cloud.

In an era of misinformation, readers need tools to critically evaluate online content. Traditional fact-checking is slow, manual, and doesn't scale to the billions of statements published daily across the web.

Debatable uses Chrome's built-in Gemini Nano AI model to analyze web content in real-time, highlighting potentially problematic statements directly on the page. Everything runs locally on your device — no API calls, no data collection, complete privacy.

• 🎯 Intelligent Sentence Extraction: Automatically identifies and extracts declarative statements from web pages
• 🤖 On-Device AI Classification: Uses Chrome's Gemini Nano model to classify statements into four categories:
  • 🔴 False: Statements that contradict well-established facts
  • 🟡 Debated: Claims with credible expert disagreement presented as fact
  • 🟠 Hyperbole: Rhetorical or promotional exaggeration
  • ⚪ Neutral: No issues detected
• 💫 Real-Time Highlighting: Sentences are highlighted inline with color-coded visual indicators
• 📊 Interactive Panel: Side panel showing all flagged statements with detailed breakdowns
• 🎨 Customizable Categories: Configure your own classification categories, colors, and definitions
• ⚡ Smart Caching: 24-hour cache with intelligent invalidation to minimize processing
• 📈 Progress Tracking: Live progress indicators with ETA estimates for batch processing
• 🔒 Privacy-First: All processing happens locally — your browsing data never leaves your device

• Glassmorphism UI: Modern, translucent legend that blends seamlessly with any webpage
• Responsive Tooltips: Hover over highlights to see classification details and confidence scores
• Keyboard Shortcuts: Navigate flagged statements quickly
• Debug Mode: Developer-friendly logging for prompt engineering and model fine-tuning

• Full Gemini Nano Integration: Replace mock classification with real AI-powered fact-checking
• User Override System: Allow users to accept/reject classifications with learning
• Dynamic Content Support: Monitor page changes and classify new content automatically
• Multi-Language Support: Privacy-preserving translation pipeline
• Browser Extension Sync: Sync user preferences and overrides across devices
• Confidence Scores: Display AI confidence levels for each classification

debatable/
├── manifest.json              # Chrome Extension manifest (MV3)
├── src/
│   ├── background.js          # Service worker: API broker, caching, message routing
│   ├── contentScript.js       # DOM extraction, highlighting, UI injection
│   ├── styles/
│   │   └── highlights.css     # Visual styling for highlights and legend
│   ├── ui/
│   │   ├── panel.html/js      # Side panel for detailed classification view
│   │   ├── popup.html/js      # Extension popup controls
│   │   └── options.html/js    # Settings page for model configuration
│   ├── classifier/            # (Reserved for future AI model integration)
│   └── util/                  # Shared utilities
└── tests/
    └── extraction.spec.txt    # Test cases for sentence extraction

• Frontend: Vanilla JavaScript, HTML5, CSS3 (no frameworks = faster load times)
• AI Model: Chrome Built-in AI (Gemini Nano via Prompt API)
• Storage: Chrome Storage API with in-memory LRU cache
• Architecture: Event-driven message passing between content script and service worker

• Chrome Canary or Dev Channel (version 128+)
• Enable Chrome Flags at chrome://flags:
  • #optimization-guide-on-device-model → Enabled BypassPerfRequirement
  • #prompt-api-for-gemini-nano → Enabled

# Clone the repository
git clone https://github.com/yourusername/debatable.git
cd debatable

# Load the extension
# 1. Open Chrome and navigate to chrome://extensions
# 2. Enable "Developer Mode" (toggle in top right)
# 3. Click "Load unpacked"
# 4. Select the project root directory

1. Navigate to any article or blog post (try news sites, Wikipedia, Medium)
2. Wait for the legend to appear in the bottom-right corner
3. Click "Panel" in the legend to open the side panel with all flagged statements
4. Hover over highlighted text to see classification details
5. Configure categories via Options (right-click extension icon → Options)

# View test cases
cat tests/extraction.spec.txt

# Manual testing checklist:
# 1. Load extension on various websites (news, blogs, social media)
# 2. Verify highlights appear correctly
# 3. Check side panel shows all classifications
# 4. Test with dynamic content (infinite scroll, SPAs)
# 5. Verify cache behavior (reload page, should be instant)

The extension leverages Chrome's built-in Gemini Nano model for on-device inference:

1. Ensure prerequisites are met (Chrome Canary/Dev 128+ with flags enabled)
2. Open extension options (right-click extension icon → Options)
3. Enable "Prompt API" checkbox
4. Model download: Gemini Nano (~1.5GB) downloads automatically on first classification
   • Requires user interaction (gesture) to trigger download
   • Progress shown in browser's download UI
   • One-time setup, persists across browser sessions

Create your own classification categories via Options page:

• Category ID: Unique identifier (e.g., satire, speculation)
• Label: Display name shown in UI
• Definition: Prompt instruction for AI model
• Colors: Background and text color (hex codes)

Example custom category:

{
  "id": "satire",
  "label": "Satire",
  "definition": "Humorous exaggeration or irony intended to mock or criticize",
  "color": "#a855f7",
  "textColor": "#ffffff"
}

• Sentence Coverage: Processes first 60 sentences by default (configurable in options, but more = slower)
• Mock Classification: Currently uses regex pattern matching; full AI integration pending
• Static Content Only: Dynamic/infinite scroll content requires page reload
• Performance: May cause lag on extremely large pages (10,000+ sentences)
• English Only: Multi-language support coming soon
• No Mobile Support: Chrome for Android doesn't yet support Gemini Nano

Problem: HTML structure makes it difficult to extract grammatically complete sentences.

Solution: Implemented a custom walker that traverses text nodes while preserving document structure, using regex to split on sentence boundaries while handling edge cases (abbreviations, quotes, lists).

Problem: Classifying hundreds of sentences would block the UI thread.

Solution: Batch processing with incremental rendering. Sentences are classified in batches of 20, with highlights applied progressively. Service worker handles heavy lifting off the main thread.

Problem: Users shouldn't have to trust a third-party API with their browsing content.

Solution: Chrome's Built-in AI (Gemini Nano) runs entirely on-device. Zero network requests, zero data collection, zero trust required.

• Best Results: Use on long-form content (articles, essays, research papers)
• Performance: Enable cache in options for instant re-classification on page revisits
• Debugging: Enable debug logging in options to see AI prompts and responses
• Custom Categories: Define domain-specific categories (e.g., "medical-claim", "legal-statement")
• Keyboard Navigation: Tab through highlights, Enter to jump to next flagged statement

We welcome contributions! This is an open-source project built for the community.

# Clone and install
git clone https://github.com/yourusername/debatable.git
cd debatable

# Make your changes
# Test locally by loading unpacked extension

# Submit PR with description of changes

Key Innovation: Leveraging Chrome's built-in AI (Gemini Nano) for completely private, on-device inference—no API keys, no cloud services, no data collection.

• Chrome Built-in AI Documentation
• Gemini Nano Prompt API Guide
• Chrome Extension Manifest V3
• Web Content Accessibility Guidelines

This project is licensed under the MIT License. Feel free to use, modify, and distribute it as per the license terms.

• Chrome team for making Built-in AI available to developers
• Open-source community for inspiration and support
• Hackathon organizers and mentors

Made with ❤️ for a more informed internet