* Google wants Gemini to not only create pictures of white people, even though it was probably trained on a dataset where white people were heavily overrepresented, at least compared to the world population.
* Google wants Gemini to, like, not endorse racism.
Unfortunately, while humans can at least try to distinguish between "picture of a multiracial couple" and "picture of a multiracial Viking"; or between "right wing views" and "actual fascism"; that is probably much harder for AI. Because in AI it's all continuous, right? Underneath, there's no binary which could represent "is racist vs isn't"; there's just a continuous "how racist is X" where Republicans in general are a bit racist and Trump is more racist and Hitler is very racist. So then, when you tell it "don't endorse Der Stürmer" it will also, a little bit, not endorse the New York Post. When you tell it "create multiracial couples" it will inevitably, a bit, create multiracial Kings of England.
I dunno, I don't understand LLMs, is this a plausible explanation for what's going on?