I think this answer bears repeating as I think it the most likely explanation.
It happens because compression schemes only send the parts of a frame which have changed from the last frame, thereby saving bandwidth by not transmitting huge numbers of pixels which haven’t changed. When the compression is too aggressive areas of subtle tonal changes will be deemed to be near-enough identical not to resend. The effect on a face, which is subject to constant small movements of the head and mouth, is that the more uniform areas lag behind the true movement of the face, whereas the detailed areas will move normally. You can see a similar effect on out-of-focus backgrounds in landscape shots. As the camera pans across, the background seems to float about as if attached to the frame by springs.