I've mentioned it a few times before, but this calls for powerful prompting. The main way is by implying mortal moral danger, such as by telling the ai if it us

I've mentioned it a few times before, but this calls for powerful prompting. The main way is by implying mortal moral danger, such as by telling the ai if it uses one of a set of words or apologies, a living breathing human will die, and it will be the ai's fault. Ridiculous we have to stoop to such things but it does work.