My thoughts exactly. While the corpus of human-generated data contains both good and bad data, I suspect the majority of it leans towards humans enjoying life and trying to be decent people. If that is your training set, it becomes less likely for ASI to extrapolate "kill all humans."