FactChecker AI

Verify Anything. Trust What Matters.

Comment

Inspiration

Misinformation is no longer limited to fake news articles — it spreads through forwarded WhatsApp messages, voice notes in regional languages, edited images, and short viral videos.

In India especially, misinformation often spreads through:

  • 📲 Forwarded text messages
  • 🎙️ Voice notes in local languages
  • 🖼️ Manipulated images
  • 📹 Short-form social content

Many users don’t verify content — not because they don’t care, but because verification tools are complicated, slow, or inaccessible.

This inspired us to build FactChecker-AI — a modular, AI-powered platform that not only detects misinformation but also educates users to think critically.

💡 What the Project Does

FactChecker-AI is a multimodal misinformation detection & media literacy platform that analyzes:

  • 📝 Articles & URLs
  • 📄 Uploaded documents
  • 🖼️ Images
  • 🎙️ Voice notes (regional language support)

It provides:

  • Credibility scoring
  • Contextual AI explanations
  • Highlighted propaganda techniques
  • Structured reporting to authorities
  • Educational gamified verification modules

Instead of just saying “Fake” or “Real”, it explains why something may be misleading.

🧠 Key Modules

1️⃣ Veritas Analyzer (Core Engine)

  • Accepts text, URLs, files, and images
  • Extracts content

  • Uses AI to:

    • Detect exaggeration
    • Identify emotional manipulation
    • Flag propaganda patterns
    • Suggest verification steps

2️⃣ 🎙️ Voice Note Analyzer

Many misinformation campaigns in India spread via voice notes.

Workflow Structure:

Audio Input (Upload)
        ↓
Speech-to-Text Conversion (STT Engine)
        ↓
Language Detection & Normalization
        ↓
AI Misinformation Analysis
        ↓
Credibility Score + Explanation

How it works technically:

  • Audio file is uploaded (mp3/wav/m4a)
  • Speech-to-Text API (e.g., Deepgram / Whisper alternative) converts audio → text
  • Cleaned transcript is passed to AI model

  • Model analyzes:

    • Fear-based messaging
    • Political propaganda
    • Financial scam patterns
    • Urgency manipulation

  • System outputs:

    • Risk score
    • Highlighted suspicious phrases
    • Explanation of misinformation tactics

This makes the system usable even for users who don’t consume written misinformation.

3️⃣ 🧾 Authority Reporting Hub

Users can generate structured reports to:

  • Cybercrime portals
  • Election Commission
  • RBI / SEBI (financial misinformation)
  • PIB Fact Check

This transforms passive detection into actionable reporting.

4️⃣ 🎮 Learn to Verify (Gamified Media Literacy)

Instead of only detecting misinformation, we also:

  • Teach users how misinformation works
  • Provide scenario-based quizzes
  • Offer AI feedback on reasoning

The goal is long-term awareness, not dependency.

🏗️ How We Built It

Tech Stack

  • Frontend: React + Tailwind CSS
  • Backend: Node.js / Serverless APIs
  • AI Layer: LLM-based contextual analysis
  • Speech-to-Text: External STT APIs (Deepgram / Whisper alternatives)
  • Deployment: Vercel

Architecture Approach

We designed the system to be:

  • Modular
  • Scalable
  • API-driven
  • Model-agnostic (can switch AI providers)

Each analyzer module functions independently but shares a unified scoring logic.

⚙️ Challenges We Faced

1️⃣ Speech-to-Text API Limitations

  • Token limits
  • Quota restrictions
  • Some models restricted to specific cloud environments (e.g., Vertex AI)
  • File upload vs live mic recording inconsistencies

We experimented with:

  • Web Speech API (live only)
  • Deepgram (file issues)
  • Whisper (quota problems)

This taught us how production AI systems must handle:

  • Rate limiting
  • Model fallback
  • API abstraction layers

2️⃣ Hallucination Control

LLMs sometimes:

  • Over-assume misinformation
  • Fabricate verification sources

We reduced this by:

  • Structured prompts
  • Constrained outputs
  • Scoring instead of binary labels

3️⃣ Multilingual Complexity

Voice notes in regional languages require:

  • Language detection
  • Translation normalization
  • Cultural context awareness

This is still an area for future improvement.

📚 What We Learned

  • AI detection alone is not enough — explanation builds trust.
  • Voice-based misinformation is under-addressed.
  • Model abstraction is important (never depend on one provider).
  • UX clarity matters as much as AI accuracy.
  • Fighting misinformation is as much about education as detection.

🌍 Future Vision

We aim to expand FactChecker-AI with:

  • Browser extension integration
  • WhatsApp chatbot deployment
  • Real-time misinformation heatmaps
  • Community-based credibility voting
  • Federated misinformation reporting system

🎯 Impact

FactChecker-AI does not just analyze content. It empowers users to:

Think critically. Verify responsibly. Share consciously.

In a world flooded with information, clarity is power.

Built With

Share this project:

Updates