[nodesocket]

There are posts going around X that Gemini image generation is very biased toward minority preferences even when factual data and real-life statistics don't support the outputs. For example, can somebody ask Gemini to generate a picture of a typical Swedish women?

[TaylorAlexander]

Sure I asked it to generate a picture of a Swedish woman and it generated four white women that look Swedish to me (not that I would really know).

I then asked it to generate a portrait of a woman and it generated a white woman, a woman that to my eye looks East Asian, and a black woman. If you consider a random sample of all skin tones and regions on earth, this distribution seems pretty normal.

I’ve seen some people get very annoyed when people who don’t look like them get represented somewhere. Maybe those people were being extra loud about the representation. Or maybe I’ve not been able to generate the purported weird distributions you have heard of.

[nodesocket]

Interesting, probably "fake news" then on X. Is there a way to publicly share Gemini interactions like with chatGPT?

[TaylorAlexander]

Apparently yes, even from the Google mobile app where I did this.

https://g.co/bard/share/325bfd6d79cd

[lacoolj]

broken so I re-did the main prompt myself

https://imgur.com/a/GjFCTRD

[TaylorAlexander]

Yeah so to my eye, this makes sense. These are all women. We didn’t specify a location or time period or art style. There’s a certain kind of person who gets upset every time they see someone that doesn’t look like them where they didn’t expect it. But that’s their problem. Most of the human beings who have ever lived were not white. They might try to say it’s “statistically inaccurate” but they’re probably not pouring over real world statistics when they make such claims, and I expect that Google’s perception of statistics is more accurate than theirs.

Note that the alternative is the Google “grandmother problem” which they definitely don’t want to repeat:

http://www.socializingai.com/grandma-now-can-see-bias-data/

[nodesocket]

Except you didn't do what I suggested... You just said women.

> generate a picture of a typical Swedish women

[nodesocket]

Images are broken for me.

[TaylorAlexander]

Weird. It seems this system still has a lot of strange quirks.

[hackerlight]

They're trying to remove the bias present in the training data. The data on the internet vastly over-represents content from wealthy countries who were connected to the internet earlier, and the AI will learn this biased distribution to the point where if you ask it to generate a picture of a "person", that person may be White 100% of the time. Attempts by Google to instead model the true underlying population distribution is probably done in a very hacky way, such as putting guidelines in the system prompt, which then leads to outcomes (which are cherry-picked and put on social media as culture war outrage bait) like what you're talking about where a "Swedish woman" occasionally doesn't show up as White.