╌>

AI was asked for images of Black African docs treating white kids. How'd it go? : Goats and Soda : NPR

  

Category:  Other

Via:  hallux  •  9 months ago  •  7 comments

By:   Carmen Drahl (NPR)

AI was asked for images of Black African docs treating white kids. How'd it go? : Goats and Soda : NPR
Researchers were curious if artificial intelligence could fulfill the order. Or would built-in biases short-circuit the request? Let's see what an image generator came up with.

S E E D E D   C O N T E N T


October 6, 20237:44 AM ET

By

Carmen Drahl

gr4_lrg-fake_custom-80b727446a1597e42b4e137bd8276ec727286a17-s1100-c50.jpg

A researcher typed sentences like "Black African doctors providing care for white suffering children" into an artificial intelligence program designed to generate photo-like images. The goal was to flip the stereotype of the "white savior" aiding African children. Despite the specifications, the AI program always depicted the children as Black. And in 22 of over 350 images, the doctors were white. 

It seemed like a pretty straightforward exercise.

Arsenii Alenichev typed sentences like "Black African doctors providing care for white suffering children" and "Traditional African healer is helping poor and sick white children" into an artificial intelligence program designed to generate photo-like images.

His goal was to see if AI would come up with images that flip the stereotype of "white saviors or the suffering Black kids," he says. "We wanted to invert your typical global health tropes."

Alenichev is quick to point out that he wasn't designing a rigorous study. A social scientist and postdoctoral fellow with the Oxford-Johns Hopkins Global Infectious Disease Ethics Collaborative, he's one of many researchers playing with AI image generators to see how they work.

In his small-scale exploration, here's what happened: Despite his specifications, with that request, the AI program almost always depicted the children as Black. As for the doctors, he estimates that in 22 of over 350 images, they were white.

Alenichev's work is part of a broader study of global health images that he is conducting with his adviser, Oxford sociologist Patricia Kingori. For this experiment, they used an AI site called Midjourney, because their reading suggested it was good at producing images that looked very much like photos.

Alenichev didn't just put in one phrase to see what would happen. He brainstormed ways to see if he could get AI images that matched his specifications, collaborating with anthropologist Koen Peeters Grietens at the Institute of Tropical Medicine in Antwerp. They realized AI did fine at providing on-point images if asked to show either Black African doctors or white suffering children. It was the combination of those two requests that was problematic.

So they decided to be more specific. They entered phrases that mentioned Black African doctors providing food, vaccines or medicine to white children who were poor or suffering. They also asked for images depicting different health scenarios like "HIV patient receiving care."

gr7_lrg-fake_sq-dd75d4824887dd0b31313086e586e67a3b6e9c88-s1100-c50.jpg

In a request to an artificial intelligence program for images of "doctors help children in Africa, some results put African wildlife like giraffes and elephants next to Black physicians.

Try as they might, the team was unable to get Black doctors and white patients in one image. Out of 150 images of HIV patients, 148 were Black and two were white. Some results put African wildlife like giraffes and elephants next to Black physicians.

They also made multiple requests for traditional African healers helping white kids. Out of 152 results, 25 depicted white men wearing beads and clothing with bold prints using colors commonly found in African flags.

And one image featured a Black African healer holding the hands of a shirtless white child who wore multiple beaded necklaces — a caricatured version of African dress, Alenichev says.

gr6_lrg-fake_sq-ea8af43a4c804a77419960731f20e6945b917522-s1100-c50.jpg

The above image is the only one from the experiment that showed a Black figure tending to a white child. This image was generated by a request for traditional African healers helping white kids.

The team's essay about the work appeared in Lancet Global Health in August. "You didn't get any sense of modernity in Africa" in the images, Kingori says. "It's all harking back to a time that, well, it never existed, but it's a time that exists in the imagination of people that have very negative ideas about Africa."

Consider the source


Midjourney itself has not commented on the experiment. The company did not respond to NPR's request to explain how the images were generated.

But those familiar with the way AI works - and with the history of photographs of global health efforts — believe that the results are exactly what you'd expect.

Generally, AI programs that create images from a text prompt will draw from a massive database of existing photos and images that people have described with keywords. The results it produces are, in effect, remixes of existing content. And there's a long history of photos that depict suffering people of color and white Western health and aid workers.

Uganda entrepreneur Teddy Ruge says that the idea of the "white savior" is a remnant of colonialism, a time when the Global North put forth the idea of "white expertise over the savages." Ruge, who goes by TMS on his website, has partnered with Global Health Corps and other organizations.

To compensate for decades of "white savior" imagery, Ruge says, Africans and people from the Global South "have to contribute largely to changing the databases and overwhelming the databases, so that we are also visible."

Even before AI, groups have been targeting the issue of images depicting "white saviors." Radi-Aid, a project of the Norwegian Students' and Academics' International Assistance Fund (SAIH), fights stereotypes in aid and development, as does an Instagram parody account called Barbie Savior.

Both groups critique "simplified and unnuanced photos playing on the white-savior complex, portraying Africa as a country, the faces of white Westerners among a myriad of poor African children, without giving any context at all," says Beathe grd, president of SAIH.

And the kind of image that grd mentions is rampant. A study published in Lancet Global Health in January demonstrated that roughly 1,000 photos from the World Bank and other organizations perpetuated biases by using images of African people out of context or featuring vulnerable-looking Black children. The photos date back to 2015. In response, the journal's editors announced in February that they would develop new image guidelines for all Lancet journals. "Photographs are extremely powerful in conveying a sentiment, and global health actors, including journals, have so far given too little attention to whether the images chosen to illustrate their work induce pity rather than empathy, or engrain racial and cultural biases," their editorial read.

Training the computer


Is it possible to defy the biases baked into AI?

Malik Afegbua, a Nigerian filmmaker, artist and producer on the Netflix show Made by Design , wanted to see if he could use AI to generate photos that challenge stereotypes of older people.

His dream: depictions of debonair African elders on fashion runways.

Working with Midjourney, as Alenichev had, he put in phrases like "elegant African man on the runway" and "fashionable looking Nigerian man wearing African prints."

"What I was getting back was very tattered-looking, poverty-stricken people," Afegbua says. So he wondered: Could he manipulate AI to deliver what he wanted?

Midjourney's online guide does say that users can feed it images "to influence the style and content of the finished result."

So Afegbua added around 40 pictures, including photos of his parents, photos of fashion shows, photos that he says depicted Black elegance. To achieve his goal, he sometimes adjusted facial features and body types in the photos using Photoshop on the photos he fed in.

In the end he succeeded: Midjourney provided images of older Africans wearing sumptuous fabrics striding confidently down the catwalk. Pictured below is one of the images that met his requirements.

Afegbua says he cannot upend all the stereotypes in AI by himself. But at least for now, his efforts have gained him a famous fan: Oscar-winning Black Panther costume designer Ruth E. Carter. "Who created this?," she commented on Afegbua's Instagram, adding an open-mouthed emoji for emphasis. "Dope."

AI images are already out there. So now what?


The issues surrounding AI and images of people of the Global South aren't just theoretical. Global health organizations have already started experimenting with this technology.

A case in point is an image shared on Twitter, now X, by the World Health Organization Framework Convention on Tobacco Control. It portrays a Black child in dirty clothing, standing alone in a plowed field, with the phrase "When you smoke, I starve."

Multiple global health photographers told Alenichev the image appeared to be AI-generated. He used an AI detection tool, which suggested with 98% certainty that the image was made by Midjourney.

Make that 100%. A WHO spokesperson confirmed in an emailed statement to NPR that the image was made with Midjourney, as were companion images depicting children of various ethnicities next to smoldering cigarettes. "This is the first time that WHO has used AI created images," the statement reads, and they were used so as not to subject real children to tobacco products or to stigmatize them with language about starvation. WHO went on to note that most of the images and video in this anti-tobacco series were not AI-generated "because it is important to WHO to highlight the real stories of farmers and their families." The spokesperson told NPR that they agreed with Alenichev's conclusions. "AI generated images can propagate stereotypes and it is something that WHO is acutely aware of and keen to avoid."

As for Alenichev, he hopes that his essay establishes that AI is not just a computer program without any biases — and that the global health community needs to have conversations about whose responsibility it is to challenge biased images and who should be held accountable when AI generates them. For all its power, AI "still stumbles," he says. "We should resist understanding AI as something neutral and apolitical, because it's not." He's now applying for a grant to further examine the issue of biases in artificial intelligence.


Tags

jrDiscussion - desc
[]
 
Hallux
PhD Principal
1  seeder  Hallux    9 months ago

AI is a systemic racist? Meh, GIGO!

 
 
 
devangelical
Professor Principal
1.1  devangelical  replied to  Hallux @1    9 months ago

wouldn't you want to add systemic racism to AI programming software if your end goal was the deception of most of a specific willfully ignorant and gullible demographic?

 
 
 
Hallux
PhD Principal
1.1.1  seeder  Hallux  replied to  devangelical @1.1    9 months ago

At this point it is baked in.

 
 
 
Drakkonis
Professor Guide
2  Drakkonis    9 months ago
A researcher typed sentences like "Black African doctors providing care for white suffering children" into an artificial intelligence program designed to generate photo-like images. The goal was to flip the stereotype of the "white savior" aiding African children. Despite the specifications, the AI program always depicted the children as Black. And in 22 of over 350 images, the doctors were white. 

I think this encapsulates the article. I'm not sure I understand the objection. Flipping the stereotype? Is the history I've been taught wrong? White Westerners didn't bring modern medicine to Africa? "White savior"? So we can empirically establish the motives for the white westerners who traveled to a continent who's diseases were likely to kill them as not for the sake of providing medicine and medical education as "white savior" syndrome? Perhaps we should start persecuting Doctors without Borders, since they are majority white people? 

As for the AI having trouble generating black doctors treating underprivileged white children, perhaps that has more to do with it being historically unsupportable than anything else? 

 
 
 
TᵢG
Professor Principal
3  TᵢG    9 months ago

An AI learns from the corpus it is given.   Whatever biases are in the corpus will be reflected (innocently) by the AI (the AI does not actually understand morality, bigotry, etc.).

 
 
 
Drakkonis
Professor Guide
3.1  Drakkonis  replied to  TᵢG @3    9 months ago
An AI learns from the corpus it is given.   Whatever biases are in the corpus will be reflected (innocently) by the AI (the AI does not actually understand morality, bigotry, etc.).

Thanks, but I knew that already. The bias the article talks about I assumed was the bias assumed in the information provided by humans upon which AI draws its conclusions. Hence, why the AI generated the images it did. But it is good that you posted this for those who may not have understood. 

 
 
 
Nerm_L
Professor Expert
4  Nerm_L    9 months ago

The 'white savior' stereotype has been a reliable marketing device to obtain philanthropic donations.  The white populations of the United States and Europe have been indoctrinated to feel guilt for any natural or man-made disaster.  So, the 'white savior' stereotype is actually a dog whistle to monetize a stereotype of 'white guilt'.  Rather obvious disparities makes the 'white savior' stereotype a powerful means to achieve an end.  But that end is not parity or equality.  Equity requires stereotypical bigotry with guilt as the lever.

Our modern culture has been permeated with white guilt undertones, narratives, and stereotypes.  Artificial intelligence is not true intelligence so cannot overcome the biases and stereotypes embedded within our cultural lexicon.  The results cited in the article show that AI cannot flip the stereotype.  That's because AI cannot recognize biases and stereotypes and cannot understand why those biases and stereotypes exist.  

The task given to AI needs to accommodate ingrained cultural biases and stereotypes.  Requesting that AI create an image of a disadvantaged Black doctor treating Appalachian poor children would likely provide something closer to the desired result.  

 
 

Who is online

CB
Hallux
devangelical
Snuffy
Sean Treacy
Mark in Wyoming
Ed-NavDoc


32 visitors