I think you might be partially wrong. For training to work, you need to feed tags and descriptions for each piece you feed, so models can weight that input to something when they're asked to generate something. The data annotation and preparation is a big part of training, probably the most important one, and it's manually done by humans, probably in third world countries like mine. It's usually the big untold part of AI, the big human data entry part of it. I'm not entirely sure how it works for videos, but probably a lot of people were paid to watch porn and annotate all the videos by timestamp, to feed along with the video for the training to happen. AI is The Turk all the way down.