So the issue is not that they don't have diverse training data, the issue is that not all things get equal representation. So their trained model will have biases to produce a white person when you ask generically for a "person". To prevent it from always spitting out a white person when someone prompts the model for a generic person, they inject additional words into the prompt, like "racially ambiguous". Therefore it occasionally encourages/forces more diversity in the results. The issue is that these models are too complex for these kinds of approaches to work seamlessly.