They were trying to combat bias from training that has hurt them in the past. The infamous, "all doctors are evidently white american men." To do that, they tried messing with the prompts that users gave, and this led to the craziness they are currently facing.
They would like to somehow get the best results of both situations. I don't think they have any method that can do that, though. You would have to know the intent of the user to know when you can go for data accurate to the training set versus data that is more accurate to possibilities. (For a fun example, Planet of the Apes have a Lincoln Memorial that I would imagine would be very hard to get from Gemini, as is.)
Make the user be the one responsible for the behavior.
Google probably has the same giant prompt but dialed to over 9000.
No. The explanation for why this has happened is because the AI was specifically instructed to do so.