Rather than training a massive deepfake detection model, we assembled existing AI models into a "paranoid forensics team" where each looks for different anomalies:
- CLIP: Semantic analysis ("Does this look like a real person?")
- Whisper: Audio-visual sync verification
- Gemini: Expert artifact detection
- Custom heuristics: Motion and audio pattern analysis
Each model votes, we fuse the results with weighted scoring.
Caught the viral airport kangaroo video and other fakes that slipped past traditional detection methods.
The philosophy: One AI can be fooled, but can you fool them all?
Rather than training a massive deepfake detection model, we assembled existing AI models into a "paranoid forensics team" where each looks for different anomalies:
- CLIP: Semantic analysis ("Does this look like a real person?") - Whisper: Audio-visual sync verification - Gemini: Expert artifact detection - Custom heuristics: Motion and audio pattern analysis
Each model votes, we fuse the results with weighted scoring.
Caught the viral airport kangaroo video and other fakes that slipped past traditional detection methods.
The philosophy: One AI can be fooled, but can you fool them all?
Demo: https://fake-check.mixpeek.com/ Code: https://github.com/mixpeek/fake-check
If you don't use "Show HN:" this is just sparkling drive-by SPAM, especially from a new-ish user.