Sigh.
The problem is that the training data is bad and they're trying to correct for it, but they're doing a terrible job, because it's super difficult to correct for bad training data.
If you ask one of these things to generate you an image of an autistic person, it generates basically the same picture over and over again of a sad, white man. Occasionally it'll pop out one woman, or one person of colour, but those results are stupid and unrepresentative. Autistic people aren't inherently sad, they're not inherently white, or inherently male. The output fails to account for any sort of diversity at all, using white men as the default. It's kind of insulting no matter how you look at it, but garbage-in-garbage-out is just as true as it ever was.
Some of our history is whitewashed. I don't know what it would generate if I asked for a person living in Palestine in the 3rd century, say. Or London during the plague, or any number of places across Europe across history, but we know that those places were not homogeneous, but our art and our media tends to reduce those places to enclaves of JUST white people, or JUST dark-skinned people, in the case of the Middle East. In reality, history shows that a lot of those places were crossroads and quite diverse.
I agree with the starting point: we should not reinforce bad data that flattens our society to just white people in all contexts, all the time. We should not have to specify the race/gender of the person just to make sure we don't get a default white person. White people aren't even the majority of people on the planet right now, it's nonsense to make that assumption in any context where the race isn't inherent in description ("programmer", "plumber", "factory worker").
But you also can't just pretend that by randomizing the makeup of any group you're going to somehow find historical accuracy. It ISN'T true that the Nazi army was racially diverse; it's absurd to claim otherwise.
In the end, it's actually just more proof that "AI" doesn't actually KNOW anything.
I don't know exactly how you fix the problem, but certainly it's more complicated than just "make sure to roll the dice every time you create an image of someone to make sure nobody is left out". That's just MORE BAD DATA, and we have enough of that.