I would be curious to know what these images would look like if it were trained on a corpus including both the dataset used AND a dataset of hand-drawn pornogra

I would be curious to know what these images would look like if it were trained on a corpus including both the dataset used AND a dataset of hand-drawn pornography, including anime and 'cartoony' porn. I'd also be curious to know how often this sort of network would bar educational content. There is an important difference between nudity and pornography. It's a distinction that I would think would be very difficult to train any automated system on, but I am prepared to be surprised.