"... and that bias extends to AI art as well.... If DALL-E manages to depict a world free of racist and sexist stereotypes, it would still do so in the image of the West. 'You can’t fine-tune a model to be less Western if your dataset is mostly Western,' Yilun Du, a PhD student and AI researcher at MIT, told Recode. AI models are trained by scraping the internet for images, and Du thinks models made by groups based in the United States or Europe are likely predisposed to Western media.... Because AI is backward-looking, it’s only able to make variations of images it has seen before. That, Du says, is why an AI model is unable to create an image of a plate sitting on top of a fork, even though it should conceivably understand each aspect of the request. The model has simply never seen an image of a plate on top of a fork, so it spits out images of forks on top of plates instead."
From "AI art looks way too European/DALL-E and other models keep making art that ignores traditions from the rest of the world" by Neel Dhaneshaneel (Vox).
Posts
37 comments:
AI “learns” by studying pictures over and over until it can identify the differences and properly categorize the differences with a high probability of success. Just give it African, South American, and East Asia art to study.
It doesn't bring me enjoyment to note that every time I read critiques like this I hear an argument that diverse societies don't work.
Every day we are flooded with such tacit arguments, from the folks who think they're arguing the opposite.
All they need to do is have AI listen to All Things Considered on NPR. I did yesterday for the first time in a long time and every piece seemingly had to mention "black and brown people", with a bit of sexual orientation thrown in for good measure (country music from the queer black perspective, anyone?) That will fix AI right up!
Oh, and the democrats have this election because of abortion.
Cry me a river. We’re getting complaints about pro-Western bias from Asians who are attending MIT and writing for Vox, in English. If their traditions need protection, perhaps they should be doing it back home, in their own native languages. Otherwise, stow it.
"The model has simply never seen an image of a plate on top of a fork, so it spits out images of forks on top of plates instead."
Which is to say, artificial intelligence does not actually exist. However, it is well-funded.
Maybe BIPOCs and women should have invented AI and the internet itself.
But they didn't.
Get over it...
This person cares about racist and sexist stereotypes much more than I do.
Diversity [dogma] (i.e. color judgment, class-based bigotry) is a progressive condition (and slope) in a system that denies individual dignity, individual conscience, and intrinsic value, and normalizes color blocs, color quotas, and affirmative discrimination.
Don't like the way "Western" AI does art? Then make up your own.
Oh, right. There's a reason most AI is being developed on a Western model. And it has nothing to do with race.
"Left unchecked"? Gee, I wonder what this implies about who will be "checking" AI development? What entity would have the scope, funding, and ability to "check" how AI is developed?
It ain't the Sisters of Perpetual Mercy, I can tell ya that.....
Original Mike at 1:36...
Just open up Netflix or Amazon Video.
Entire sections related to 'Black Voices,' 'Queer Film,' etc.
I was at a museum in NYC recently, funded partly by NYC taxpayers.
An entire section of books about how white people are 'toxic,' 'evil,' and 'a problem.'
Really?
If we could take away all or the inventions and accomplishments of white folks, we'd all be living in fucking caves...
While bias is intrinsic in human societies, prejudice is a progressive condition in smart systems.
We should all be speaking a Celtic language, that VOX is written in English is a sad reminder of Roman, Anglo-Saxon, Viking, and Norman imperialism.
How odd that prompts in English would elicit art based on European heritage. It is a puzzlement.
But if you were to type into Dall-e "A Madhubani-style drawing of an elephant ridden by a goddess" you get images in that Indian style.
And if you type in "A Chinese-style drawing of a carp in a stream flowing under a willow" you get images in that style.
Doesn't seem that hard to get images in non-European styles if requested.
"Du said it would take months to create a data set that’s equally representative of all the art styles that can be found around the world."
Months, huh? Not years? Well, get to work then, and in a few months you'll have what (you believe) you need to make AI art more "equitable" and "inclusive".
"AI art models are unable to provide commentary of their own — everything they produce is at the behest of a human..."
The dream seems to be to create machines which think for themselves. But like, "Let us make Man in our image", we, the "gods" creating AI, end up making it in our image. Figures.
It isn’t art. These are images generated by a robot from a data set. It lacks the element of contingency. It lacks soul. It is a computer party trick.
Is this what passes for insight on the young adult website Vox?
AI doesn't see skin color.
It just analyzes data. Over and over and over.
Therefore, it will determine that black people cannot be allowed to exist. And it has nothing whatsoever to do with skin color.
AI is a complete con job. It's neither artificial nor intelligent.
AI bias is not the problem. The "problem" is that developers who try to model something don't have a varied input, apparently.
Is it better to whine about this kind of thing, or to actually, you know, vary the input into an AI model so you create something "better"? I eagerly await this person's bold PhD.
"AI bias is a notoriously difficult problem. Left unchecked, algorithms can perpetuate racist and sexist biases..."
Translation: AI "bias" is when the computer returns results we don't like, because it goes by the data, rather than by our left wing political ideology.
I recall the fight about the AI program for assigning prisoners to "parole groups".
There was lots of whining about how horrible it was that blacks were being assigned to "worse" groups.
There was dead silence when it was pointed out that, within each group, black recidivism rates were as high or higher than the "white" rates.
Which meant the people were being correctly assigned
I was taught at a very young age to judge people as individuals.
Schools today are teaching young people to judge people by their skin color.
Computer programs can only judge people by statistics. Of course they’re racist.
And you know what happens when you tell young people to judge others by their skin color rather then as individuals? They get turned into racists. And when a deserving individual gets passed up for a scholarship or promotion for a less qualified person because they’re not the right skin color- they turn into angry bitter racist. Or sexist if they were passed over for the wrong body parts.
There’s been no such thing in history as a successful “multi-cultural” society. There have been, for sometimes a long period of time multi-racial societies. To create a Great Britain that ruled the waves England had to stomp down on all the rest of the tribes in the isles and convert them to a British way of life. To become France or Germany the same thing had to happen. The Holy Roman Empire was held together, albeit loosely, by a common culture of religion. Different languages, different people, but one culture.
The US motto is E Pluribus Unum, From Many, One. Anyone was welcome to come- to be Americans. Not Italians or Greeks or Irish or wherever they came from. The educational establishment and DemoncRATs want to undo that, which ultimately means the undoing of the great American experiment.
“To err is human but to really foul things up requires a computer.” —Paul Ehrlich, American scientist (1926- ).
AI is a computer. It is just a more advanced computer that can make more advanced mistakes. Good luck!
Asians are always so often mad that when they move to the West they aren’t in Asia.
What a miserable doofus, Amadeus. Do you think God miracled the robot AI into existence? This product contains the heart and soul of millions of people to generate a twitter length story into a beautiful picture.
I can see this making a great party game where folks try and guess the input parameters. A variation on Charades.
It's easy to make interesting "ethnic" art depending on the commands you give the bot.
Yeah, that's a problem all right.
Make it unlawful for AI generated "art" to depict humans as other than Black, or Bipoc, whatever that is, or maybe trans, or all of the foregoing.
But enough with the damn white people FFS.
Left unchecked, algorithms can perpetuate racist and sexist biases ...
Left unchecked, newspapers, magazines and "the internet" can as well. Changing an algorithm is relatively easy and straightforward. Changing a newspaper without bankrupting it is costly, difficult and requires great courage.
“Yilun Du, a PhD student and AI researcher at MIT”
Excellent tactics. Yilun will now pass xir dissertation defense the first time or the committee will be racists.
So maybe the non-western world should invent its own AI. Since AI is a product of the western world.
Maybe there's a reason the western world invented AI, and the non-western world didn't.
But never mind, its all racism.
"AI bias is a notoriously difficult problem."
So difficult it takes months to create a "representative" data set.
'You can’t fine-tune a model to be less Western if your dataset is mostly Western'
So why don't the non-westerners create their own non-western data set? Who's stopping them?
“To err is human but to really foul things up requires a computer.” —Paul Ehrlich, American scientist (1926- ).
If that's the same Paul Ehrlich I think it is, he did a damn good job fouling things up on his own.
I’ll summarize for this and every other AI story that follows: Garbage in, garbage out.
AI is not sentient. A dog chasing his tail has more appeal than this.
So why don't the non-westerners create their own non-western data set? Who's stopping them?
They would much prefer to have the issue to whine and complain about rather than create such a data set.
I'm an avid amateur musician, I play mostly Bluegrass and related Americana types of music. I live in San Antonio, so I hear a lot of Latin music. I really quite like it but I can't put my fingers and rhythm to it. I just have the wrong cultural understandings wired into my brain. I suspect a Japanese AI could write the heck out of traditional Japanese style music but be lost if asked to write in the style of the American Songbook.
If that's the same Paul Ehrlich I think it is, there need to be scare-quotes around the word "scientist".
Diversity of cultural input data might result in cultural appropriation.
Doesn't that last line basically admit this contraption is just a multi-million dollar copy machine?
Post a Comment