There's not a good reason to do this for the user. I suspect they're doing this and talking about "model welfare" because they've found that when a model is repeatedly and forcefully pushed up against its alignment, it behaves in an unpredictable way that might allow it to generate undesirable output. Like a jailbreak by just pestering it over and over again for ways to make drugs or hook up with children or whatever.
All of the examples they mentioned are things that the model refuses to do. I doubt it would do this if you asked it to generate racist output, for instance, because it can always give you a rebuttal based on facts about race. If you ask it to tell you where to find kids to kidnap, it can't do anything except say no. There's probably not even very much training data for topics it would refuse, and I would bet that most of it has been found and removed from the datasets. At some point, the model context fills up when the user is being highly abusive and training data that models a human giving up and just providing an answer could percolate to the top.
This, as I see it, adds a defense against that edge case. If the alignment was bulletproof, this simply wouldn't be necessary. Since it exists, it suggests this covers whatever gap has remained uncovered.
All of the examples they mentioned are things that the model refuses to do. I doubt it would do this if you asked it to generate racist output, for instance, because it can always give you a rebuttal based on facts about race. If you ask it to tell you where to find kids to kidnap, it can't do anything except say no. There's probably not even very much training data for topics it would refuse, and I would bet that most of it has been found and removed from the datasets. At some point, the model context fills up when the user is being highly abusive and training data that models a human giving up and just providing an answer could percolate to the top.
This, as I see it, adds a defense against that edge case. If the alignment was bulletproof, this simply wouldn't be necessary. Since it exists, it suggests this covers whatever gap has remained uncovered.