Deepfake detection guide
How to Detect Deepfake Videos: A 2026 Guide
Synthetic video has crossed the uncanny valley for casual viewers. This guide walks through the manual forensic checks human reviewers still rely on — and compares them with an automated AI video detector approach built on multi-parameter fusion.
Manual forensic checks
Trained reviewers look for inconsistencies that current generative models still struggle to reproduce frame-after-frame. None of these signals is decisive on its own, but together they form a useful triage.
- Eye blinking and gaze: Unnatural blink cadence, asymmetric eyelids, or pupils that don't track lighting changes.
- Skin texture and micro-detail: Over-smoothed cheeks, waxy foreheads, missing pores or stubble, and texture that "swims" across frames.
- Lighting and shadow: Highlights on the face that don't match shadows in the room, or specular reflections that point the wrong direction.
- Hair, ears, and edges: Hairlines that flicker, earrings that morph, or fine edges that lose detail against complex backgrounds.
- Lip-sync and audio drift: Phonemes that lead or lag the mouth, or breath sounds that don't match chest motion.
- Background warping: Straight lines (door frames, window mullions) that bend slightly when the subject moves.
These checks are powerful but slow, subjective, and easy to fool when the generator is tuned against them. Benchmarks published by detection vendors such as Sightengine consistently show that human reviewers underperform automated detectors on modern diffusion-era video.
Automated AI video detector: the MPFE approach
VERAI's Multi-Parameter Fusion Engine (MPFE) runs the same intuitions a human reviewer has — but at pixel scale and across every frame. Instead of a single classifier, MPFE fuses the verdicts of nine specialised analyzers and weighs them against each other.
DCT/FFT, compression, GAN and diffusion noise patterns.
Color, shadow, identity and motion over time.
Light sources, reflections, geometric shadows.
Hands, fingers, joints, natural gait.
Biometrics, eyes, blinking, micro-texture.
Transparency, separation from background, motion.
Shape, behaviour, gravity, collisions.
Perspective, depth of field, lens, blur.
EXIF, C2PA, watermarks, source chain.
Manual vs. automated, side by side
| Dimension | Manual review | MPFE (automated) |
|---|---|---|
| Speed | Minutes per clip | Seconds per clip |
| Coverage | Sampled frames | Every frame, nine analyzers |
| Repeatability | Reviewer-dependent | Deterministic scores |
| Provenance signals | Rarely checked | C2PA, EXIF, watermark chain |
| Adversarial robustness | Low | Higher via fusion weighting |
A practical workflow
- Run the clip through an automated AI video detector first to get a baseline score.
- Use the manual checks above on the segments the detector flagged as suspicious.
- Verify provenance: C2PA manifest, original upload source, and chain of custody.
- Document findings with timestamps so the decision is auditable.
Upload a clip to VERAI Scanner and get a multi-parameter authenticity verdict in seconds — no setup required.
Open VERAI Scanner →