How VeriVidDetects AI Content
A transparent look at our five-engine forensic pipeline. We believe you should understand exactly how we analyze your content — no black boxes.
Results are probabilistic, not absolute proof
No detection system achieves 100% accuracy. VeriVid results are risk indicators with confidence scores. They should be considered alongside context, source provenance, and corroborating evidence before drawing legal or editorial conclusions.
Five-Engine Forensic Analysis
Each engine examines a different forensic layer. Signals are combined into a weighted risk score with a calibrated confidence percentile.
👁️01Visual ForensicsPrimary Signal · 40% weight
Primary Signal · 40% weight
- Pixel-level inconsistencies and diffusion rendering artifacts
- GAN-specific noise patterns unique to generative models
- Facial boundary irregularities and blending edges
- Temporal consistency across multiple extracted frames
- Morph seams and spatial frequency anomalies
SightEngine AI + custom ensemble classifiers
🎙️02Audio IntegritySupporting Signal · 25% weight
Supporting Signal · 25% weight
- Synthetic speech patterns and voice-cloning signatures
- Spectral frequency anomalies in audio waveforms
- Lip-sync timing mismatches with the video track
- Background noise consistency across the recording
- Prosody and phoneme cadence inconsistency signals
Spectral analysis + NLP prosody models
🗂️03Metadata IntelligenceSupporting Signal · 20% weight
Supporting Signal · 20% weight
- Container format and codec fingerprint analysis
- Encoding timestamps and software watermarks
- Metadata field stripping (common in manipulated files)
- Re-encoding signatures from post-processing pipelines
- Platform-specific metadata consistency validation
FFprobe-based extraction and rule engine
⏱️04Temporal ConsistencySupporting Signal · 10% weight
Supporting Signal · 10% weight
- Frame-to-frame motion vector analysis
- Unnatural stabilization or optical flow artifacts
- Flicker patterns inconsistent with natural capture
- Edit boundary detection and cut anomalies
OpenCV optical flow + custom temporal scoring
🤖05AI Risk AggregatorFinal Scoring · Ensemble
Final Scoring · Ensemble
- Weighted ensemble of all signal scores
- Confidence calibration against benchmark datasets
- INCONCLUSIVE detection for low-signal inputs
- Evidence summary generation for report attachment
HuggingFace models + custom scoring layer
How to Interpret Results
No significant AI manipulation signals. Content appears authentic.
Some signals detected. Manual review and corroborating verification recommended.
Strong manipulation signals. Treat with caution and verify source independently.
Insufficient signal data — poor quality, very short clip, or novel AI not in training.
Limitations & Accuracy
✅ What VeriVid Detects Well
- ·GAN and diffusion-generated faces/bodies
- ·Face-swap and deepfake manipulation
- ·Synthetic voice and audio cloning
- ·Lip-sync manipulation (lip-dub attacks)
- ·Re-encoded and post-processed video
⚠️ Where Detection May Fail
- ·Heavily compressed content (low bitrate)
- ·Novel AI models not yet in training data
- ·Very short clips with limited frame data
- ·Screen-recordings of AI-generated content
- ·Traditional (non-AI) video editing
Accuracy Benchmarks (internal testing)
~85%
High-quality deepfakes
~90%
AI-generated imagery
~75%
Compressed social clips
Accuracy varies based on source quality, AI model type, and compression level.