How VeriVidDetects AI Content
A transparent look at how we analyze your content: independent AI-detection models on the frames and audio, C2PA provenance verification, and an agreement-gated confidence layer. No black boxes — and no overclaiming.
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.
The Detection Layers
Each layer contributes a different kind of evidence. Model probabilities drive the score; provenance can confirm it; and confidence is gated on independent signals agreeing.
👁️01Visual AI-Generation DetectionPrimary scored signal
Primary scored signal
- An AI-generation model scores each sampled frame (0–100% likelihood)
- Detects diffusion/GAN rendering artifacts and morph seams
- Pixel-based — robust to re-encoding and platform compression
- Both the mean and the peak frame are considered, so localized fakes aren't averaged away
SightEngine genai model (Pro can add Reality Defender as a second opinion)
🎙️02Audio Synthetic-Speech AnalysisScored when an audio track is present
Scored when an audio track is present
- Estimates the likelihood that speech is synthetic / voice-cloned
- Excluded from the score when there is no audio track — never treated as 'low risk'
- Combined with the visual signal (visual-dominant) when both are available
Audio detection model via the provider layer
🔏03C2PA Content Credentials (Provenance)Deterministic · near-proof when present
Deterministic · near-proof when present
- Reads cryptographically-signed C2PA manifests embedded in the media
- A valid manifest declaring AI origin is verifiable proof — it overrides to HIGH RISK
- A signed camera/human capture corroborates authenticity
- Absence of a manifest proves nothing (most social video is stripped) and is reported as such
C2PA / Content Authenticity Initiative verification
🧮04Agreement-Gated CalibrationFinal scoring layer
Final scoring layer
- A single detector caps at MEDIUM confidence — one model's confident mistake never reads as certainty
- HIGH confidence requires two independent detectors to agree
- Strong disagreement returns a contested REVIEW at LOW confidence, not a falsely-precise number
- INCONCLUSIVE when frames/audio can't be obtained; metadata & heuristic flags are shown as context but never change the score
Calibrated, evidence-based scoring
How to Interpret Results
No significant AI-generation signal. Content appears authentic.
Some signal, or independent detectors disagree. 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
Why we don't publish a single accuracy number
A headline figure like “98% accurate” is misleading: it comes from controlled benchmarks that don't reflect novel generators, heavy social-media compression, or the base-rate problem (when few videos are fake, even a strong detector produces many false alarms). Instead of a vanity number, every result carries a calibrated confidence level, only reaches HIGH when independent detectors agree, and returns INCONCLUSIVE rather than guessing. When verifiable C2PA provenance is present, we report that as near-proof.