Is this video real?

Four models check the face, the voice and the lip movement of a talking-head video, and look for signs of an AI video generator. They tell you what they found and how far to trust it.

How it works

1. Give it a video

Upload a clip of one person talking, or record one with your camera or microphone. A voice recording or a picture works too.

2. Four models check it

One looks at the face, one looks for a generated video, one listens to the voice, and one checks whether the lips move with the words. The whole video is checked, three seconds at a time.

3. You get a plain answer

Signs of manipulation, can't tell, or no signs found, with each model's score and a timeline showing which moment stood out.

How well it works

Measured on clips the models never trained on, with the same thresholds the site uses.

Fakes of the kinds it learned
72%

caught, on people it never saw: face swaps, redrawn mouths and cloned voices.

Fakes from newer tools and other sources
4% to 71%

caught, depending on the source. The voice model catches most fully generated videos; the face model misses them.

Real videos wrongly flagged
3% to 6%

depending on where the video comes from; about 5 in 100 on average.

All test fakes together
28%

of 6,677 fake clips from four datasets caught, while 6% of 4,481 real clips were wrongly flagged.

Your video stays private

Your video is checked on our computer and deleted before you see the answer. Nothing is kept or shared.

Limits

  • "No signs found" is not proof a video is real. On footage unlike the training data, the checks miss many fakes.
  • About 5 in 100 real clips get flagged. That rate was measured on four research datasets; for anything else it is an estimate.
  • It needs one face and speech. Other clips get "can't check".
All the limits