I don't think that's really true - most smart people are capable of considering both of those values (amongst many, many other values) and come up with a reason

I don't think that's really true - most smart people are capable of considering both of those values (amongst many, many other values) and come up with a reasonable balance where, for example a prompt for a founding father would be white, but maybe a prompt of a hero from the American Revolution was not. Just as an example. But Gemini is no where near smart enough to achieve that kind of nuance, so instead you can kind of have to tune it with pretty blunt pre-text. So, the engineering team saw previous complaints about image generators providing too "stereotypical" images in some contexts and pushed the generator more towards returning "diverse" results. But Gemini isn't a smart human, it's a blunt machine and so it doesn't know in which contexts "diversity" is appropriate. I think the underlying issue here is that until these AI tools genuinely are smarter than smart humans (and probably even after), the second they get released they're going to be under massive scrutiny. And no person, let alone some crappy AI is going to give perfect answers to 1,000,000 users asking questions attempting to trip it up. The solution to this is obvious though - you accept it's not perfect! The problem is that's not an acceptable solution for either Google's brand or Google's shareprice.