It’s still machine learning, you need labelled data to teach the algorithm what is what. It’s exactly the same thing as with captcha, random people labelling pictures for self driving cars. I doubt they were watching videos in real time, they were probably going through footage afterwards and marking errors and such.