… for example, I wanted to generate an avatar for myself; to that end, I want it to be representative of me. I had a rather difficult time with this; even explicit prompts of "use this skin color" with variations of the word "white" (ivory, fair, etc.) got me output of a black person with dreads. I can't use this result: at best it feels inauthentic, at worst, appropriation.

I appreciate the apparent diversity in its output when not otherwise prompted. But like, if I have a specific goal in mind, and I've included specifics in the prompt…

(And to be clear, I have managed to generate images of white people on occasion, typically when not requesting specifics; it seems like if you can get it to start with that, it's much better then at subsequent prompts. Modifications, however, it seems to struggle on. Modifications in general seem to be a struggle. Sometimes, it works great, other times, endless "I can't…")

We project everything onto AIs. Unbias in LLMs doesn't exist.

Not all humans though.