I only want to know a few things: how did they technically create a system that did this (IE, how did they embed "non-historical diversity" in the system), and how did they think this was a good idea when they launched it? It's hard to believe they simply didn't notice this during testing. One imagines they took steps to avoid the "black people gorilla problem", got this system as a result, and launched it intentiona…
The problem with it is that training on model output is a well known way to screw up ML models. Notice how a lot of the generated images of diverse people have a very specific plastic/shiny look to them. Meanwhile in the few cases where people got Gemini to draw an ordinary European/American woman, the results are photorealistic. That smells of training the model on its own output.