How InstaTruth Works
Our advanced multi-agent fact-checking system combines DistilBERT NLP classification with comprehensive web verification using Google Custom Search and DeepSeek AI analysis.
Audio Transcription & Analysis
When you submit a video link, our system downloads the content using yt-dlp and uses OpenAI Whisper to transcribe any spoken content into text with high accuracy.
The transcribed text is then cleaned and prepared for analysis by our BERT-based natural language processing pipeline.
Transcription Result:
"Studies show that this new treatment has a 90% success rate in clinical trials and could be available as early as next month..."
Sources Found:
- Clinical trial data from National Health Institute (Phase 2 results only)
- No sources confirm availability timeline
Multi-Step Claim Extraction & Verification
DeepSeek AI first extracts individual verifiable factual claims from the transcription. Each claim is then independently researched using Google Custom Search API to find credible supporting or contradicting sources.
Each claim undergoes detailed fact-checking with source credibility weighting (.gov, .edu, major news outlets), providing granular analysis with confidence scores and evidence summaries.
DistilBERT NLP Classification
Simultaneously, our DistilBERT transformer model (110M parameters) processes the transcribed text to generate semantic embeddings, which are classified using a pre-trained Random Forest model for initial credibility assessment.
This NLP analysis provides probability scores standardized to P(true) with confidence thresholds, offering an independent linguistic-based credibility assessment.
Classification Results:
Overall Truth Score
Dynamic Score Aggregation & Report
The system combines DistilBERT and web verification scores using dynamic confidence-based weighting. High-confidence web fact-checking (when available) takes precedence, while NLP analysis provides fallback assessment.
Final reports include individual claim breakdowns, source credibility analysis, confidence scores for each component, and comprehensive summaries with final verdicts: "real", "fake", or "inconclusive".