My understanding, however fuzzy, is that all the safety/politeness tuning results in models that are at times less likely to give accurate responses. That said,

My understanding, however fuzzy, is that all the safety/politeness tuning results in models that are at times less likely to give accurate responses. That said, I suspect that either way both types of models largely give similar answers for soft questions aside from those politeness and safety things