It’s probably not the base model. It’s probably the hidden prompt augmentation rules that everyone has been scrambling to add to these models.

Gemini, like Chat GPT, it it likely appending simple prompts with more detail behind the scenes so results come out more varied. For example, since the models regurgitate the content people promote on the internet, if you searched for “attractive person” last year, you’d probably get a white person 95% of the time. When something is over represented in mainstream media, gen AI will reflect that back.

Now, before something goes into the gen AI black box, it is secretly given racial and photo style modifiers so users get more diverse people and image styles unless the user manually specifies what they want to see.

Generally this works well, but it can backfire hilariously if you ask for it to draw a group of people who are famous to having a certain ethnicity.