Earlier quoted context omitted.
The only problems I see are that someone could potentially upload a lower resolution version of the video a moderator has seen and then perhaps bypass the content id system (I don't know how tolerant they tend to be for situations as that). There's also the aspect of if we make it search out for distinct frames (like "scene change" type things), there's the possibility that some safe things could become auto-moderate…
Even with lower resolutions, fingerprinting (on a frame level) should still work relatively well, shouldn't it? If you get a certain level of confidence on each frame - if you have 80% of the frames with 80% confidence to be the same as $someVideo, that ought to be good enough overall to not get false positives but catch repeat views - how many 30s videos will randomly share 80% of their frames with a 90s video of an…
If a powerful actor wants the 30s video removed, and knows it would be removed if it shared 80% of its frames with a 90s video of an animal being abused, then it would not be technically difficult to engineer such a 90s video.