Comment by jholdn

This is somewhat to credit AI model design. I think this is how you'd want such models to behave with esoteric subject matter. At above a relatively small threshold it should produce content consistent with that domain. Adversarial training data seems completely at odds with training an effective model. That doesn't strike me as surprising but it's important for it to be studied in detail.