I am rather skeptical about PhotoDNA. If it is an effectively method for video filtering then why is youtube using very expensive machine learning, which has a

I am rather skeptical about PhotoDNA. If it is an effectively method for video filtering then why is youtube using very expensive machine learning, which has a high maintaince and operation costs, compared to just simply hashing the video frames. There is also similar problem with spam where spammers send email with images in order to fool the spam filter. If the algorithms in PhotoDNA would be effective then the problem of spam images would be a fairly solved problems, but what I keep hearing is that the only effective tool is machine learning.