��о����ʦ�ҷ��黨��Ƶ

A recent study led by a ��о����ʦ�ҷ��黨��Ƶ researcher examined how AI reflects stereotypes and how users can disrupt these biases to create more authentic, affirming images of Black girls.

Jennifer D. Turner, professor of literacy education, collaborated with her niece, independent researcher Amani R. Danridge-Boston, to explore ways to use OpenAI’s chatbot, ChatGPT, to create visual images of Black girls that honor their humanity and dignity. Specifically, Turner wanted to create an image for an upcoming keynote address that would depict a Black girl navigating multiple forms of discrimination based on race, gender and age, while shaping her own destiny. The study was published in the .

The researchers began their investigation with an understanding that AI is not neutral but instead reflects biases and stereotypes that are embedded in its training data and rooted in media and the wider society. For Black women and girls, these include the Jezebel stereotype, which hypersexualizes and objectifies them, and the Strong Black Woman stereotype, which expects them to sacrifice themselves, deny their pain and constantly demonstrate strength.

“Given the stereotypes that we know are embedded within the training data, we need to disrupt them and really refine the prompts to reduce those kinds of harms,” said Turner.

The researchers led ChatGPT through four rounds of image generation, creating multiple images each time and refining their prompts after each round. To help them analyze the images, they used an intersectional multimodal framework Turner developed that draws upon the work of Black feminist scholars. The framework guided them to examine factors including how the Black girl character occupied space within the image and how her identity, voice and agency were represented. Some specific elements they examined included layout and color palette, as well as the character’s skin color, hairstyle, clothing, facial expression, point of view, and size and scale. 

Although ChatGPT did present a culturally appropriate range of natural hairstyles, several harmful stereotypes of Black women and girls emerged in the images. In early rounds, ChatGPT generated images of young Black women in various states of undress, reflecting the Jezebel stereotype. In later rounds, the researchers realized they needed to specify articles of clothing in their prompts to disrupt this stereotype. 

“It was surprising to me, the extent to which we had to work to reduce and disrupt that bias, because those biases were embedded in the visuals right in front of our faces, from the very beginning,” said Turner. “We really had to be intentional and specific with the language in the prompting to get it to generate more humanizing representations of Black girls.”

As they continued to refine their prompts, they found that the clothing the Black girls wore in some of the later images also reproduced harmful stereotypes. For example, they asked ChatGPT to depict a Black girl standing in an intersection of several roads, which symbolized multiple forms of oppression. The chatbot produced several images of young Black women wearing dresses made out of the same types of colorful wavy lines as the roads. The researchers found that this reflected the Strong Black Woman stereotype, as the young women were literally wearing the oppressive forces that surrounded them.

“We don’t want to take on the oppressions of the world. We’re trying to dismantle them, not take them on as clothing for ourselves to wear,” said Turner. “We believe Black girls should be clothed in joy, liberation, happiness and authenticity.”

During the fourth round, ChatGPT finally generated an image that the researchers felt represented the diversity of Black girls’ experiences of joy, pain, strength, vulnerability, oppression and freedom. The final image had a more subtle color palette than previous images, and the girl’s face was in profile, in contrast to earlier images which showed a young woman with her back to the viewer. Although white lines symbolizing multiple oppressions were encroaching on the girl’s body, her face remained untouched and surrounded by a bright light as she gazed out past the image’s borders. This final image can be viewed in the Journal of Visual Literacy article.

For others who want to create humanizing images of Black girls and women using AI, the researchers recommend four strategies: acknowledge that AI is not neutral and is trained on data that reflects societal biases, use clear and specific instructions in prompts to disrupt the bias inherent in AI, use a Black female-centered visual framework to help analyze stereotypes in AI-generated images of Black girls, and keep experimenting and refining the prompts to create more nuanced, humanizing images.

Turner also encourages educators to help students think critically about AI. For example, students might consider whether AI-generated images are empowering or harmful for Black girls and how the images could better represent Black girls’ humanity.

“Many times, young people think of AI as being an expert, but they should not be afraid to speak back to AI,” Turner said. “You can collaborate with AI, but having a high level of criticality is crucial.”

In her next steps, Turner plans to collaborate with Black girls to learn more about how they are using AI to generate images and how they have used prompts to create the kinds of images they want to see.

Turner remains optimistic about the potential for users to recalibrate AI to reduce bias and generate images that represent the complexity of Black girls’ lives and experiences.

“AI is a tool,” Turner said. “We can use specific language and our own knowledge about authenticity and representation and work alongside Black girls and Black women to create visual representations that are humanizing and affirming.”

Photo courtesy iStock/Mariia Vitkovska

The images included here were generated as part of the researchers’ study but were not published in the original journal article.