If I understand your position: you don't have any better ideas, and you're okay with using ML even if it means deleting evidence of human rights violations. Is that accurate?
I'm not particularly sympathetic to the notion that YouTube should be obligated, either legally or ethically, to host content in perpetuity. OTOH, if rapid filtering results in such content being taken down so quickly that even archivists and reporters don't have a reasonable chance at identifying and saving it, then I do see that as a potential ethical (but not legal) dilemma, and an interesting technical challenge.
My initial thought upon reading the title was that this story--such as it is--was more a commentary on the unintended consequences of the censorship of terrorist-related propaganda videos. From that perspective, I don't think the distinction between ML and human filters is that important; the more interesting issue seems to be the scope and creep of censorship and how we can expect tech companies to draw equitable lines within modern political cultures that suffer from hyper-partisan political discourse. It's not hard to imagine a situation where simply permitting any atrocity videos will be seen as partisan--supporting one side or another by promoting particular narratives.
The alternatives are
1. Algorithms, which always going to fail in human rights type "edge cases" like the ones described in the article, and which really shouldn't be seen as edge cases IMO, or
2. Bury your head in the sand and allow terrible content to exist.
Do you have any better ideas for how to deal with processing traumatic content at scale?