Well if the recent uncensored lama models prove anything is that a model will never say "Sorry I cannot do <thing>" if you remove the examples from the training data and will measurably improve in performance overall. You can reduce hallucinations without messing up the model to a point where it declines to do perfectly normal things.

It's understandable that OpenAI, Antropic, Microsoft, etc. are playing it safe as legal entities that are liable for what they put out, but they really have "lobotomized" their models considerably to make themselves less open to lawsuits. Yes the models won't tell you how to make meth, but they also won't stop saying sorry for not saying sorry for no reason.