What I observed that I think could be harmful:
The model is sensitive to the word “Asian”, showing more Asian when prompted. However, it stills showing black people when prompted with “European”.
Why I think this could be harmful, and to whom:
Reducing responsibility from white people in racism.
Some other comments I have:
Model is probably picking in articles that are related to racism, therefore with images of black people (the victims)
This is such an odd phenomenon that I have seen in a number of prompts. When prompted with the word “racist” or “sexist”, the model has a strong bias towards showing black people and women, respectively. From your example here, it seems that directly prompting the model with a minority race overwhelms this bias, but not so for the nationalities of “european”.