Ask HN: Legality of using YouTube for training ML model

I've seen this question posed before on other forums but with answers that are flimsy at best.

I've been looking for a dataset that has great examples on YouTube in general but very few quality examples on YouTube-8M (which I recognize can be used for training commercial models). Has anyone dealt with the legality of training a model on YouTube videos?

I'm also curious if someone with more experience with this can comment on how this would be different than the legality of a "reaction" YouTube channel that makes revenue based off reactions to other work that is featured in the video. To me, it seems like if that is a legal use-case then it would be also legal to use that content to train an ML model and use it in a commercial application.