The title of the blog post seems to have changed since submission. It's worth noting (and the article tries to argue this if you read below the examples) that

The title of the blog post seems to have changed since submission. It's worth noting (and the article tries to argue this if you read below the examples) that while a number of right-leaning (or outright right-wing) talking heads took this opportunity to talk about White Genocide and "erasing Europeans from their own history" and other white supremacist talking points, as expected, this is most likely just a very stupid overcorrection to compensate for the general racism of the training data. If you go back a few years, there are numerous articles about how AI perpetuates racist and sexist stereotypes[0] because the training data contains real world biases[1] (e.g. men and women or white people and Black people being overrepresented in different professions - "show me a doctor" always turning up a white guy, "show me a cleaner" always turning up a Latina woman, "show me a criminal" always turning up a Black man, etc[2]). There was also an article on here last year about how historical "selfies" always looked American, effectively overriding the cultural norms of what they were supposed to represent. It's clear this was some hamfisted attempt to avoid bad press that hilariously backfired. Depicting German soldiers from WW2 as brown people in Nazi uniforms is not woke revisionism, it borders on satire. I don't know how you can seriously believe that depicting Nazis as ethnically diverse is part of some leftist agenda but here we are. Also, the reason it refuses to generate images of "white British royalty" seems honest enough: as that narrows it down to a very specific group of real world individuals (i.e. it would likely result in pictures of King Charles or Queen Elisabeth and other members of the royal family) whereas "Black British royalty" likely doesn't match any known real world individuals and thus makes it past the filter. The nonsense pictures of the founding fathers are likely a mix of the two problems: it refusing to generate images of known individuals but also the founding fathers being a group of people and thus subject to the diversity overcompensation. [0]: And let's not forget the gem of a faux pas when someone found out that Google's image classifier routinely labeled photographs of Black people as "gorillas" rather than recognizing them as humans. [1]: And specifically biases from data typically gathered in the US. A model trained on data from Nigeria would still be biased even if it routinely generated images of Black people in suits for any given profession. "Bias" here is an objective claim, not a moral one. [2]: Again, this is true regardless of your personal politics. A model fed with present day data only for example would end up generating pictures of men when asked to show you a programmer in the 1960s despite the gender bias in that profession being largely the inverse at that time to what it is today.