Deepfakes can swap faces, but they don't understand physics. Light doesn't reflect correctly. Shadows fall the wrong way. Faces move in ways real muscles can't. These physics violations are often the clearest tells that something is fake. This guide catalogs the ways deepfakes violate physical reality.
How to Use This Guide
Each section covers a category of physics violation:
- The Violation: What physical law is being broken
- What It Looks Like: Visual symptoms
- Why It Happens: Technical explanation
- Detection Method: How to spot it
- Real-World Example: Practical illustration
Category 1: Lighting Violations
Violation: Shadow Direction Mismatch
The Violation: Shadows on the face fall in a different direction than shadows in the scene.
What It Looks Like:
- Face has shadows on the left; scene has light from the left
- Nose shadow points up while other shadows point down
- Face appears lit from above while room is side-lit
Why It Happens: The source face was recorded under different lighting. The algorithm transfers the face but not the lighting. Shadow direction from the original recording persists.
Detection Method:
- Identify the main light source in the scene
- Check if shadows on the face match that direction
- Compare shadow direction on face vs. other objects in frame
Real-World Example:
"The interview was shot with a window on the right. But the shadow under the person's nose pointed left—toward the window. That's physically impossible."
Violation: Specular Highlight Inconsistency
The Violation: Shiny spots on skin don't match the light source positions in the scene.
What It Looks Like:
- Forehead shine suggests light from above-left; lamps are above-right
- Highlights on cheeks don't correspond to any visible light source
- Specular highlights remain static when face moves
Why It Happens: Specular highlights are highly view-dependent. The source footage's highlights don't match the target scene's geometry. The algorithm doesn't recompute highlights based on scene lighting.
Detection Method:
- Look for shiny spots on forehead, nose, cheeks
- Trace imaginary lines from those spots to potential light sources
- Check if highlights move correctly as face moves
Violation: Color Temperature Clash
The Violation: The face has different color temperature than the rest of the scene.
What It Looks Like:
- Face looks warm (yellowish) in a cool-lit (bluish) room
- Face appears daylight-balanced while scene is tungsten-lit
- Skin tone doesn't match visible lighting
Why It Happens: Color temperature is baked into the source footage. Basic color matching can shift overall tone, but precise temperature matching across lighting variations is difficult.
Detection Method:
- Compare the white balance of the face vs. other skin in the scene
- Check if face color is consistent with the visible light sources
- Look at ear and neck color (often not processed) vs. face
Category 2: Reflection Violations
Violation: Eye Reflection Mismatch
The Violation: Reflections visible in the eyes don't match what should be in front of the person.
What It Looks Like:
- Eyes show a window reflection; there's no window in that direction
- Ring light reflection in eyes; no ring light in scene
- Different reflections in left and right eye when they should match
Why It Happens: Eye reflections from the source footage are transferred. These reflections show the source's environment, not the target scene. Most algorithms don't attempt to recompute eye reflections.
Detection Method:
- Zoom in on the eyes
- Identify shapes reflected in the cornea
- Check if those shapes exist in the scene from that angle
Real-World Example:
"In the video, you can see a square window reflected in their eyes. The scene was shot outdoors. There was no window anywhere near them."
Violation: Missing or Wrong Environmental Reflections
The Violation: Reflective surfaces near the face don't reflect correctly.
What It Looks Like:
- Glasses reflect a different room
- Shiny jewelry shows wrong reflections
- Water or mirrors show the original face, not the deepfake
Why It Happens: Reflections in nearby objects aren't processed. They show whatever was in the source footage or target footage—not a consistent synthetic scene.
Detection Method:
- Check any reflective surface near the face
- Compare the reflection to what should be visible
- Look for reflections of the face itself in nearby surfaces
Category 3: Anatomical Violations
Violation: Impossible Facial Movements
The Violation: The face moves in ways human facial muscles can't achieve.
What It Looks Like:
- Lips stretch wider than humanly possible
- Eyes move independently in unnatural ways
- Facial features slide rather than deform
- Expressions that don't correspond to any real emotion
Why It Happens: The algorithm maps one face's geometry onto another. When the source face makes an expression that requires different muscle geometry than the target face has, impossible distortions result.
Detection Method:
- Watch for expressions that feel "wrong" without being obviously fake
- Look for asymmetry that doesn't match typical human asymmetry
- Observe if features slide across the face rather than deform naturally
Violation: Neck-Head Disconnect
The Violation: The head moves in ways that don't match neck position or body posture.
What It Looks Like:
- Head rotation that would require the neck to bend impossibly
- Head appears to float on top of body
- Rotation axis seems wrong (head pivots around wrong point)
- Head tilt doesn't cause corresponding neck skin movement
Why It Happens: The algorithm processes the face, not the neck. When the source head position differs significantly from target body position, the connection becomes anatomically incorrect.
Detection Method:
- Watch the neck during head movements
- Check if the rotation axis makes anatomical sense
- Look for skin stretching or bunching that should occur but doesn't
Violation: Blinking Anomalies
The Violation: Blink patterns don't follow normal human physiology.
What It Looks Like:
- No blinking for extended periods (30+ seconds)
- Mechanical, perfectly rhythmic blinking
- Asymmetric blinking that doesn't match individual's normal pattern
- Blinks that are too fast or too slow
Why It Happens: Early deepfakes famously forgot to include blinking. Modern systems include it but may not capture individual blink patterns. Source and target blink patterns may conflict.
Detection Method:
- Count blinks over 30-60 seconds
- Normal adults blink 15-20 times per minute
- Watch for unnatural regularity or irregularity
Real-World Example:
"The person in the video went 45 seconds without blinking. That's not normal. Humans blink every 3-4 seconds on average. Either they were a statue or something was fake."
Category 4: Physics of Motion
Violation: Inertia Violations
The Violation: Face motion doesn't follow laws of inertia.
What It Looks Like:
- Head starts and stops movement too abruptly
- No acceleration/deceleration curve
- Cheeks don't jiggle slightly during head stops
- Hair doesn't continue moving after head stops
Why It Happens: Per-frame processing doesn't model physical momentum. Each frame is computed independently, so there's no "memory" of previous motion that would create natural deceleration.
Detection Method:
- Watch for abrupt motion starts and stops
- Look at soft tissue (cheeks, under-chin area) during motion
- Observe if hair momentum matches head movement
Violation: Gravity Inconsistencies
The Violation: Elements don't fall or hang correctly.
What It Looks Like:
- Hair doesn't fall naturally when head tilts
- Tears don't roll down in straight lines
- Loose skin doesn't respond to gravity during head position changes
Why It Happens: The algorithm doesn't simulate physics. It maps textures and shapes but doesn't understand that gravity pulls things downward.
Detection Method:
- Watch hair behavior during head tilts
- Observe how skin responds when head position changes
- Check if any hanging elements (earrings, etc.) behave correctly
Violation: Collision Failures
The Violation: The face passes through objects it should collide with.
What It Looks Like:
- Hand passes through face without deformation
- Glasses clip into face geometry
- Hair passes through shoulder without bouncing
Why It Happens: The algorithm has no concept of solid objects. It processes pixels, not physical surfaces. Occlusions are handled by layering, not by simulating physical interaction.
Detection Method:
- Watch moments when objects touch the face
- Look for skin deformation that should occur but doesn't
- Observe if objects maintain impossible positions relative to the face
Category 5: Temporal Physics
Violation: Cause-Effect Reversal
The Violation: Effects appear before or without their causes.
What It Looks Like:
- Smile begins before the emotional trigger
- Eyes show reaction before the stimulus
- Speaking begins before audio starts
- Expression changes with no apparent reason
Why It Happens: Source and target footage may have different timing. Emotional expressions from the source may not align with events in the target scene.
Detection Method:
- Watch for emotional responses that don't match context
- Check audio-visual synchronization
- Observe if reactions precede their causes
Violation: Temporal Drift
The Violation: Physical properties gradually change in impossible ways.
What It Looks Like:
- Face proportions slowly shift over time
- Skin color gradually changes throughout video
- Features drift position without corresponding movement
- The person looks subtly different at the end than the beginning
Why It Happens: Each frame is processed somewhat independently. Small variations accumulate over time, causing gradual drift that wouldn't occur in reality where faces maintain consistent proportions.
Detection Method:
- Screenshot the face at beginning, middle, and end
- Compare proportions and colors
- Look for changes that have no physical explanation
Category 6: Material Physics
Violation: Skin Behavior
The Violation: Skin doesn't behave like actual skin tissue.
What It Looks Like:
- Skin stretches uniformly like rubber rather than realistically
- No subsurface scattering (skin appears flat, not translucent)
- Pores don't scale correctly with face movement
- Wrinkles appear painted on rather than as actual skin folds
Why It Happens: Real skin is a complex layered material with specific optical properties. The algorithm processes it as a texture, not as a physical material with thickness, translucency, and mechanical properties.
Detection Method:
- Look at skin during expression changes
- Check if wrinkles look like surface texture or actual folds
- Observe translucency in thin areas (ears, around nostrils)
Violation: Fabric Interaction
The Violation: Clothing and face don't interact correctly.
What It Looks Like:
- Collar doesn't move with chin
- Tie or scarf clips through neck
- Glasses frames pass through ears
- Hats seem to float above or cut into head
Why It Happens: Face processing stops at the face boundary. Objects that should physically interact with the face are processed separately (or not at all), creating discontinuities.
Detection Method:
- Watch clothing near face during movement
- Check if accessories maintain physical contact
- Observe transitions between face and non-face areas
Quick Reference: Physics Violations
| Category | Key Violations | Primary Detection Method |
|---|---|---|
| Lighting | Shadow direction, highlights, color temperature | Compare to scene lighting |
| Reflection | Eye reflections, surface reflections | Check what should be reflected |
| Anatomy | Impossible movements, neck disconnect, blinking | Watch for unnatural motion |
| Motion | Inertia, gravity, collisions | Observe physics of movement |
| Temporal | Cause-effect, drift | Compare across time |
| Material | Skin, fabric interaction | Watch physical interactions |
Summary
Deepfakes process pixels but don't understand physics. Lighting doesn't recompute for new scenes. Reflections show the wrong environment. Faces move in anatomically impossible ways. Motion violates inertia. Materials behave like textures, not physical substances.
These physics violations are often the most reliable tells because they require understanding that current algorithms don't have. A convincing texture can't compensate for a shadow that falls the wrong way.
When evaluating suspected deepfakes, ask: "Does this follow the laws of physics?" If it doesn't, you've found your answer.
Related Topics
- Why Do Deepfakes Still Look Wrong? Common Failure Modes – Common failure modes
- How Can You Tell If a Video Is a Deepfake? – Detection guide
- Why Do Deepfake Expressions Look Wrong? – Motion and expression issues
- Why Does My Deepfake Face Look Wrong? – Facial detail problems
- What Can't Deepfakes Do Yet? – Current technology limits