Comment by jsheard

DALL-E/Bings silent injection of keywords to try to de-bias the model produces funny results since the model got better at producing text, because the injected keywords tend to appear as text in the generated image, so you may end up with a character randomly saying "ethnically ambiguous" in a speech bubble or something else along those lines.

e.g. https://twitter.com/StyledApe/status/1709728954993557932

https://twitter.com/DerekPutin/status/1728928441507189069