How to Spot a Deepfake Video
A field guide to catching face swaps, lip-sync attacks, and full-synthesis video by eye — plus the limits of human detection and when to reach for the automated detector.
If you've landed here because you think a video you're watching might be a deepfake, the short answer is: your eyes are probably not going to settle this for you, but they can narrow the odds before you reach for a detector.
Here's what to look at, in order of reliability.
The five tells that still work in 2026
1. The jaw-line and hairline
Face-swap models operate per-frame. Where the transplanted face meets the original head, there's often a faint blending ring — especially visible at the jaw and along the forehead. Pause the video. Zoom in. Look for:
- A slight texture shift between cheek and neck
- Hair that seems to float slightly in front of or behind where it should
- Skin color that doesn't quite match the neck
2. Head rotation past ~45°
Most face-swap pipelines are trained primarily on front-facing reference footage. When the target turns their head substantially toward profile, the model's coverage breaks down. Watch for:
- Noticeable loss of detail when the face turns
- Nose or ear geometry that "snaps" between two positions rather than rotating smoothly
- Glasses that visibly detach from the face during rotation
Attackers know this. They avoid profile shots. A video that's suspiciously always front-on is a soft flag by itself.
3. Lip-sync timing
Lip-sync deepfakes pair real video with new audio + regenerated mouth region. The sync is usually close but often slightly off: phonemes lead mouth shapes by 40–90ms rather than the naturally-recorded ~20ms. To notice this:
- Watch the mouth while listening to the audio — does it feel slightly "ventriloquist"?
- Look for mouth transitions that seem to complete just after a sound has already been made
4. Blink rhythm
Humans blink 15–20 times per minute in normal conditions, with chaotic spacing. Early deepfakes under-blinked visibly; modern systems have mostly solved this but can still produce too-regular rhythms. If blinks appear mechanically spaced or oddly timed with speech, note it.
5. The obvious background stuff
Shadows from a light source the scene doesn't have. Reflections in glasses that show a different scene. Text in the background that's garbled or subtly wrong. Full-synthesis video (Sora, Runway, Veo) still fails some environmental consistency checks even when the main subject looks great.
When to stop looking and use the detector
If you've watched the video three times and any of the following are true, you've hit the limit of human inspection:
- The clip is <10 seconds (not enough time for tells to accumulate)
- The compression is heavy (TikTok, Reels, screen-recorded from a phone)
- The subject is someone whose natural video you haven't seen much of (no mental baseline)
- The stakes justify a second opinion
At that point, use our free video deepfake detector. It runs the video and audio tracks through DETECT-3B Omni in parallel and returns:
- A verdict and confidence score for each track
- A list of the specific cues the model flagged
- A generator match — which family of deepfake pipeline the video most closely resembles
For lip-sync attacks specifically, this is the only reliable defense — a dual-track detector catches the cloned audio while the visual track passes inspection.
What doesn't work
Things that seem like they should be tells but aren't reliable in 2026:
- Finger count. Frontier video models got this right; stop relying on it.
- Eye reflections. Used to reveal synthetic faces. Most generators now match environment light plausibly.
- "It looks fake." Intuition is an unreliable signal. If you can't say why the video looks wrong, write it off.