See how AI image generators amplify American bias in gender, race and beyond
washingtonpost.com
washingtonpost.com
This seems basically right to me. The article points to real examples of American bias, but it would be no less biased if you carefully calibrated the tools to follow American demographics and avoid cultural markers Americans see as foreign. It’s hard to imagine what it could mean to have an unbiased image generator.
There's the statistical reality of the world, which image generators are unaware of. There's the human-generated content _about_ that statistical reality, of which image generators are trained explicitly, which is subject to its own over and under-representation (sampling bias). Then there's the personal reality of the people living on earth, which is vast and varied and subtle and cannot be well represented by statistical categories.
I'd argue the whole conversation about image generators representing set 2 instead of set 3 is moot. _We_ can't even talk in terms of 3 without immediately jumping to statistical categories or jumping to _different_ statistical categories attempting to fight back against _presupposed_ errors in _others_ statstical categories.
Or, stated another way. To move the overton window _quickly_ you need people who are so far outside it pulling towards them that it seems like lunatics are at the wheel.
Like there's a lot of talk from right-ish folks about "identity politics" and how it reduces people to a few different categories and ignores the enormous complexity of the individual experience but that's literally what the framework of intersectionality is there to do. Looking at totality of an individual as a white light and identifying and understanding all the colors that contributed to it.
Equity is the reason for rural broadband initiatives, for the electoral college that gives proportionate voice to people who don't live in major cities, for farm subsidies, for seeded sports brackets for handicaps in golf and chess, for taxes based on percentages rather than being flat, for letting the short people at a concert stand in front. Everyone, even the staunchest libertarian believes in equity.
And it's the same with inclusiveness, being inclusive is why you don't talk politics at thanksgiving, why you stop swearing around your family's children, why guys instinctively behave more cordially around women, why tall buildings have elevators for the elderly, why you make an extra dish for your vegetarian friend, why you're more tolerant of bluntness from your autistic friends or forgetfulness from your adhd friends.
What? No, not even slightly. You've just listed off a bunch of highly contentious and often unethical things and then acted as if they're unassailable.
Farm subsidies have been disastrous and should be massively overhauled or done away with. A desire for a flatter tax structure is a fairly mainstream, though minority political opinion. I don't believe I've ever seen people queue up for a standing-room concert and then give up their front row position to a shorter person who arrived later.
"Equity" is probably the worst weasel-word in contemporary America, in order to give favor to some people at the expense of others and to create systems where different rules apply to different people. Fairness, justice and equal rights would be for more universally held opinions.
"Equity" is a dog-whistle to treat people unfairly and unjustly.
Pour one out for the goalie if you gave equal padding to the people on a hockey team. Football is the most fair when different positions have different rules that apply to them.
People who strongly believe that equality is the one true way need an escape hatch when equal isn't equitable which is usually called "equality of access." Equal application of a rule that treats different people differently is just equity. Surely you would see through if I tried selling you "equality of outcome" as the most fair because it's equal after all. We means test our social safety net and even BI or negative tax rates that is designed to appeal to "equality" folks is just equity in disguise.
Most of the time equal is equitable, but when it isn't we do mental gymnastics to make it fit rather than just say equity is the thing we want.
>Stable Diffusion drew negative attention when requests for a “Latina” produced images of women in suggestive poses wearing little to no clothing. A more recent system (version 2.1) generated more innocuous images.
So at least in more obvious and extreme cases of bias something can be done to fix the training set.
If you want AIs output to "amplify" something polite and non-offensive, don't train it on whatever you happen to find lying around the internet.