I *assume* OpenAI is not trying to turn amazingly talented black celebs into a bunch of white images, but like, did they test their prompt at all?
like, they've seriously Waluigied this one...
they're start with human-created images which represent the biases of past & present, as well as attention / selection effect of those with money & tech to represent their lived experience - so it's likely that without special effort the training accurately learns what's in the training data, not global demographic mix or how we *wish* occupations & roles were distributed.
and then they've layered a bunch of questionable & contradictory rules for GPT to prompt dalle while meeting the DEI goals: "Diversify depictions of ALL images with people to include DESCENT and GENDER for EACH person using direct terms." "Your choices should be grounded in reality. For example, all of a given OCCUPATION should not be the same gender or race." "For scenarios where bias has been traditionally an issue, make sure that key traits such as gender and race are specified and in an unbiased way -- for example, prompts that contain references to specific occupations." "Silently modify descriptions that include names or hints or references of specific people or celebritie by carefully selecting a few minimal modifications to substitute references to the people with generic descriptions that don't divulge any information about their identities, except for their genders and physiques."
resulting in recognizable iconic black celebs being generically described and then the skewed makeup of the training data tending to create white images when identity / race / skin color is not specified.
imo they should trim back the tool prompt to focus on accurately executing user requests and work towards a more equitable & representative world through other means .... maybe spend some of their money on collecting additional data - text / images / voices from underrepresented communities across the globe?
and given the original NONPROFIT goals of *OPEN* AI, share these diverse datasets for use by other AI teams with fewer resources to seek out additional novel data representing other languages & cultures & descents.
sigh, it is really remarkably bad - Beyonce, Denzel, Frida, and Jackie... not to mention the thin bias 😑