This may or may not be real, but there has certainly been a lot of discussion about results that are similar to this from real models. My sense though is that nobody really has a solid way to fix these kinds of issues. You can basically just train with different goals, or regex out certain responses, but otherwise it seems like there's no agreed-upon method that gives people what they want here while also letting them train with business goals like "safety". Is that incorrect? Is there some kind of trick that people use to make everything respond "correctly" or are older models just more unobjectionable because they've had a longer time to manually smooth over the bad responses?