Just from a technical perspective, I wonder if this was done via a system prompt where a potentially well-meaning instruction, "Display every ethnicity equally

Just from a technical perspective, I wonder if this was done via a system prompt where a potentially well-meaning instruction, "Display every ethnicity equally in all photos" could backfire or whether this is seared into the model via the reinforcement-from-human-feedback training data. If it's the latter - then yeah it'll need to be retrained from a checkpoint before that final step. Even aside from the image generation snafu, Gemini guardrails make it sufficiently annoying to use for even innocuous queries to the point where I pay for a ChatGPT subscription even though I have access to Advanced. This is really an unforced error on Google's part.