5 Signs a Video Has Been Deepfaked (+ How to Check)

Video is harder to fake convincingly than a still image — there are simply more frames to keep consistent, and motion, lighting and audio all have to agree across time. That’s good news for you: deepfaked video still leaves tells, especially in movement and sound. Here are the five most reliable signs, how to check each one, and why no single sign is a verdict on its own.

How to inspect any clip: download it (screen-recording re-compresses and hides artifacts), play at 0.25× speed, and step frame by frame. Full screen, brightness up. Most tells live in the transitions between frames, not in any single paused image.

1. Edges that shimmer or “breathe”

Watch the boundary where the face meets hair, ears, glasses or the background. Face-swap models blend a generated face onto a real head every frame, and that seam often flickers, warps or shifts slightly between frames — a subtle “breathing” around the jaw, hairline or neck. Pause on a fast head-turn or when a hand passes in front of the face: that’s where blending breaks down.

2. Lip-sync that drifts

Audio and mouth shapes can fall fractionally out of step, especially on plosive sounds (p, b, m) where the lips should fully close. Watch the mouth with the sound off, then with it on. Warning signs:

• Lips that don’t quite close on “m”/“b” sounds.
• Teeth and the inside of the mouth that look smeared, static or oddly uniform.
• A voice that’s almost synced but consistently a hair early or late.

3. Lighting and colour that don’t agree

Real light obeys one physics across the whole scene. In composited or generated video the face may not match its environment:

• Skin lit warm while the room is cool (or vice-versa).
• Shadows on the face pointing a different way than shadows in the scene.
• A face that stays evenly lit while the person moves through light and shade.
• A faint colour “halo” or mismatched grain/noise where the face was inserted.

4. Unnatural motion, blinking and micro-expressions

Faces do tiny things constantly — blink at a natural rate, make micro-expressions, move the head with subtle jitter. Synthetic faces often get this subtly wrong:

• Blink rate too low (early deepfakes barely blinked) or mechanically regular.
• Eyes that don’t track where the person is “looking,” or a fixed, glassy gaze.
• Too-smooth or too-still motion — an eerie steadiness, or a face that seems to float slightly relative to the head.
• Frozen forehead/ears while the mouth moves a lot.

5. Missing provenance and suspicious metadata

Step away from the pixels and check the file and its origin:

• A clip that claims to be a real recording but carries no camera metadata, or metadata that contradicts the story (wrong date, wrong device, editing software).
• No credible source — it appears only on anonymous or fringe accounts, with no outlet carrying it.
• Conversely, valid C2PA Content Credentials are evidence for authenticity.

Bonus: listen to the audio

Voice cloning has its own artifacts. With headphones, listen for a flat, “even” delivery with little breath, odd pacing or emphasis, a subtle metallic ring, or background noise that cuts unnaturally between words. Audio-only scams (the “family member in trouble” call) rely on you not listening this closely.

What a forensic tool does automatically

You can catch a lot by eye, but software goes where you can’t:

• Optical-flow & temporal consistency — checks that objects move the way real objects do, frame to frame.
• Frame-by-frame neural face analysis — flags the subtle artifacts blending leaves behind.
• Frequency analysis — the statistical fingerprint of generative models, invisible to the eye.
• Metadata & provenance — EXIF, encoding consistency and C2PA in one pass.

For the image-specific version of these checks, see how to tell if an image is AI-generated; for the background on how these fakes are built, see what is a deepfake.

Quick checklist

1. Download, slow to 0.25×, step frame by frame.
2. Watch the face edges on head-turns and occlusions.
3. Check lip-sync on p/b/m sounds; inspect teeth/mouth interior.
4. Compare lighting/shadow on the face vs the scene.
5. Watch blinking, gaze and micro-motion.
6. Check the source and metadata; listen to the audio on headphones.

The bottom line

No single sign is proof — real clips have compression, motion blur and odd crops too. The reliable approach is to combine the signals, which is exactly what a forensic tool does: it fuses several independent checks into one verdict instead of betting on a single tell. Verifyco runs that fusion on your iPhone in seconds, fully offline — see on-device verification, explained.