Deepfake detection tutorial
Learn to identify AI-generated and manipulated video through visual inspection, audio analysis, and forensic tools.
Hands-on tutorial · 20 min
Before you start
This tutorial teaches practical deepfake detection skills. You'll learn what to look for visually, how to analyze audio, and when to use forensic tools.
Important: Visual and audio inspection alone cannot reliably detect high-quality deepfakes. These skills should be combined with forensic tools (like ClipForensics) and source verification for comprehensive analysis.
Part 1
Visual inspection
Train your eye to spot common deepfake artifacts.
Check faces at frame edges
Face-swap deepfakes often show artifacts at the boundary between the swapped face and original background. Look for blending seams, color mismatches, and resolution differences along the jawline, hairline, and ears.
Examine eyes and teeth
AI-generated faces frequently show eye reflection inconsistencies (different reflections in each eye), irregular pupil shapes, and unnaturally uniform teeth. Zoom in and scrub frame by frame.
Watch hands and fingers
Current AI generators struggle with hands. Look for extra fingers, impossible joint positions, fingers merging or disappearing, and morphing appendages across frames.
Check text and symbols
AI-generated video often produces garbled text on signs, shirts, or screens. Real video reproduces text clearly; synthetic video frequently gets letters wrong or produces nonsensical characters.
Observe background stability
In AI-generated video, background objects may gradually morph, drift, or change shape across frames. Pause on a background detail and advance frame by frame to check consistency.
Part 2
Audio analysis
Listen for signs of synthetic speech and audio manipulation.
Listen for breathing
Natural speech includes breathing between phrases. Many TTS and voice cloning systems omit or produce unnatural breathing patterns. Listen for unusually clean transitions between sentences.
Check speech rhythm
Synthetic speech may have unnatural prosody — the rhythm, stress, and intonation of speech. It may sound "too smooth" or have odd emphasis on certain words.
Evaluate lip-sync
Slow the video to 0.5x or 0.25x speed and watch if lip movements precisely match the spoken words. Lip-sync manipulation creates subtle but detectable timing misalignments.
Listen for environment
The acoustic environment (room reverb, background noise) should match the visual setting. A person appearing to be outdoors should not sound like they are in a recording studio.
Part 3
Metadata and technical checks
Look beyond pixels and audio at the file itself.
Check file metadata: Right-click the file and examine properties. Look for encoder software, creation timestamps, and camera model information. Absence of camera metadata is suspicious (but not conclusive).
Use forensic tools: Upload the video to ClipForensics for automated analysis across 15 forensic modules. The tool examines dimensions that are impossible to check manually — spectral frequency patterns, compression history, quantization fingerprints, and biological motion statistics.
Review the evidence timeline: If forensic analysis flags anomalies, check whether they're clustered in specific segments or distributed throughout. Clustered anomalies may indicate a spliced segment; distributed anomalies suggest full-video generation.
Part 4
Putting it all together
Effective deepfake detection combines all three approaches:
- Visual inspection — quick triage to identify obvious artifacts.
- Audio analysis — check speech naturalness and synchronization.
- Technical analysis — use forensic tools for comprehensive, automated examination.
- Source verification — who shared it, when, and why?
- Contextual assessment — does the content make sense in the real world?
No single step is sufficient. The combination provides far stronger conclusions than any individual check.
For a structured process, use our video authentication checklist.