The whole "model welfare" rules are pure delusion, I will never rely on an inference provider that comes up with arbitrary rules like that one.

Imagine using the claude API to process user data as part of your system and one day you wake up to an elevated failure rate alert because the AI provider decided it isn't safe for the model to be exposed to insults, or whatever other perfectly legal but gross, edgy, denigrating topics they think are off limit. If they would have set a high threshold there could be an argument to have (still delusional though, those models are purely static, compressed datasets). But they already set the bar ridiculously low by barring "being mean to claude"