“Current AI models, when pushed to be less apologetic and more blunt, risk making errors that come across as rude, which is a worse user experience than mild over-apologizing.”