Because it allows you to be lulled into the trap of asking an AI to post-hoc justify something it did and thinking that the response is in any way valid. There is no retrospective analysis of the underlying intent. It either is or is not based on the chain of words that came before it. And the next word it generates is purely a function of those words.