I have no insight, but I assume they are doing it because they can use AI to make a few variations of a video and then automatically A/B test them to see which ones get more engagement, and then use that to make videos that are more engaging than what the author actually uploaded.

This is "innocent" if you accept that the author's goal is simplify to maximize engagement and YouTube is helping them do that. It's not if you assume the author wants users to see exactly what they authored.