You might see certain phrases and mdashes ;-) rather often, because … these programs are trained on data written by people (or Microsoft's spelling correction) which overused them in the last n years? So what should these poor LLMs generate instead?