[Alex-C137]

“They are so racialised. They should never even let those be published because it’s like the worst stereotypes about Africa, or India, or you name it,” said Alenichev.

This is a problem and always has been with AI, people have been saying this for at least a decade at this point. Type in "photorealistic picture of child in refugee camp" in any AI that lets you.

[guerrilla]

Isn't it just probabilistic? That's who's most likely in poor and/or in refugee camps.

[jncfhnb]

50,000 ft priors on probability don’t extend to good perspectives on specific settings

[guerrilla]

I don't know what this means.

[jncfhnb]

Citing that the average poor person is, say, African, is not useful when discussing a specific poor area in, say, Asia.

[guerrilla]

I still don't understand what you're trying to say.

What I was saying is if you ask an LLM to generate an image of a poor person, it makes sense they'd be brown or black because if you were to randomly pick actual poor people from Earth, the chances are very high it'd be a brown or black person. In this case, it's just accurate representation.

[jncfhnb]

And the problem is that often times we are not talking about random poor people, but rather, specific populations of poor people whose demographics and other traits do not match the highest level average. So the LLM is entirely wrong.

[whimsicalism]

what would your ideal racial probability distribution for someone in a refugee camp be?