Isn't this just due to more compression? What's stopping someone from turning down the bitrate and re-encoding a video into different formats a few times to kill the quality (which will still look fine on a 6 inch phone in portrait orientation)?
That's one of the effects we model in fact, "social media laundering" is the term of art.
Various detectors are more or less thwarted by it. It actually surprised me how strong the artifacts from some GANs are - they can survive several passes of re-encoding, but accuracy does suffer.
But I still need the raw 720/1080 stream for training.
