That being said, I don't believe they should be doing anything like this without the creator's explicit consent. I do personally think there's probably a good use case for machine learning / neural network tech applied to the clean up of low-quality sources (for better transcoding that doesn't accumulate errors & therefore wastes bitrate), in the same way that RTX Video Super Resolution can do some impressive deblocking & upscaling magic[2] on Windows. But clearly they are completely missing the mark with whatever experiment they were running there.
[1] https://www.ynetnews.com/tech-and-digital/article/bj1qbwcklg
[2] compare https://i.imgur.com/U6vzssS.png & https://i.imgur.com/x63o8WQ.jpeg (upscaled 360p)
YouTube is not applying any "face filters" or anything of the sort. They did however experiment with AI upscaling the entire image which is giving the classic "bad upscale" smeary look.
Like I said, I think that's still bad and they should have never done it without the clear explicit consent of the creator. But that is, IMO, very different and considerably less bad than changing someone's face specifically.
