But I dont thing feeding n's responses into n+1's training data is safe, even if its vetted by humans.