The argument being that there is not very much video footage of people turning their heads, therefore not enough data to train deep fake videos / filters.
Videos are still very much in the baby phase. There are way, way easier tells when a video has been faked. We're talking about images. Turned head images are very much in scope.