Social Memberships and Identity Representation in “Text-to-Image” Artificial Intelligence Programs

Artificial Intelligence (AI) programs are subject to the same biases and prejudices as their human engineers – the surface of a problem we are just beginning to scratch. The intersectional reality of how oppression pervasively advantages some and disadvantages others, persists even in the innovative spaces made possible by collaboration with AI programs. But we cannot address what we cannot yet see. Here, we describe comparative analysis and art projects that explore how narrative designers, writers


INTRODUCTION
As multimodal creators, we explore, challenge, and reimagine social constructs using generative Artificial Intelligence (AI) programs through a humanist and creative lens. We are also members of a new cohort of creators who have benefited from the reduced levels of prerequisite knowledge needed to use these technologies. In the course of our respective work, we have encountered abundant examples of the ways pervasive biases are transmitted or transmuted visually within these programs. As Kate Crawford evocatively questions about AI: "whose interests does it serve, and who bears the greatest risk of harm?" (Crawford 2021). Our resulting observations and explorations are visual responses to this question, demonstrating how some aspects of ethnicity, sexual orientation, and gender identity are represented and misrepresented within the AI text-to-image programs, NightCafe and Midjourney.
With the pace of AI advancement, we acknowledge that our work is a snapshot in this rapidly evolving timeline, serving as additional examples in the landscape of interdisciplinary conversations about algorithmic bias (Buolamwini 2018), the various ways this bias privileges whiteness (Noble 2018), and how computer systems throughout history have reinforced supremacist norms and hierarchies (Hicks 2019).
As marginalized communities are most at risk of harm from the scalability of transmitted bias and topdown implementation of AI systems (Reventlow 2021), our participation as impacted creators is essential. As a form of conscientious resistance and action, we document how bias impacts our humancomputer collaborations in order to inform the ongoing development of prosocial and humane generative AI technologies.

STUDENT PROJECTS AT SOUTHERN ILLINOIS UNIVERSITY CARBONDALE
As part of SIUC's Master of Fine Arts program, four graduate students began using generative AI during a narrative workshop course that explored multimodal design and gamification. The students encountered embedded social messages that communicated certain identities as "other" (Staszak 2008) from the mainstream visuals prioritized in the AI programs. In response, they documented the transmission of bias they were witnessing and exercised resistance against these dominant modes (Madaio et al. 2021) by implementing strategies to bypass the biases upheld by the programs.

A machine learning approach to black bodies & beauty standards
Black bodies, identities, and aesthetic representation are often distorted, devalued, and portrayed as a kind of "otherness." For centuries, Eurocentric ideologies of beauty have been upheld as the desirable norm, and these ideologies are seen in AI as well. The generative AI programs' interpretations of beauty default to Eurocentric features. This default has significant implications for Black communities, and specifically, for Black women. Academics such as Bell Hooks, Richard Dyer, and Shawna Hudson, believe the way Black women are represented in the media affects how they are seen by others and how those others interact with them. Bell Hooks argues that the media does not accurately represent Black women, yet it "determines how blackness and people are seen and how other groups will respond to us based on their relation to these constructed images" (Hooks 1992). Hudson furthered that media curated images of Black women come from dominant racial, gender, and class ideologies: "These stereotypes simultaneously reflect and distort both the ways in which [B]lack women view themselves (individually and collectively) and the ways in which they are viewed by others" (Hudson 1998, p.49).
Especially relevant to Dajonea Robinson's work is Dyer's observation that Black female representation is informed by studies of whiteness.
"The only way to see the structures, tropes, and perceptual habits of whiteness, is when non-white (and above all, black) people are also represented" (Dyer 1997).
The images Robinson produced show this effect, as they simultaneously reflect and distort the ways in which Black women view themselves and the ways in which they are viewed by others.
For example, Figure 1 was created using the generative AI program, NightCafe. The initial prompt used was "Black fairy with glass wings, beautiful pastel colors". The fairy is detailed in many ways, from her two sets of wings to the intricacies of her facial features. The wings have intricate ruffles and the face is symmetrical with all features detailed and present. The word "Black" only appears in the fairy's hair and wings, as opposed to being a marker of identity. Robinson changed the prompt to include "African American" and Figure 2 demonstrates several inadequacies. There are less fine details in this fairy's facial features and outward appearance. The skin tone and curly hair are representative, but overall the image is a disservice. The AI could not conceptualize a Black fairy and instead composed a shell of a character compared to its Eurocentric default counterpart. This phenomenon has also been documented by others creating AI-generated images from the biases and limitations of the whitedominated datasets used to "train" AI (Salvaggio 2019).
With that in mind, it is important to note that AI can be programmed to generate ideas in a variety of ways. Some systems mimic human creativity, while others use more algorithmic approaches such as Generative Adversarial Networks (GANs). The methods used to collect data for NightCafe and Midjourney are unknown because the companies have not released this information to the public. However, it can be inferred that the AI uses a combination of sensors, cameras, the internet, and other inputs to generate its art.
As a result of this information, the learned Eurocentric defaults can become so entrenched that they can be layered over Black subjects in a type of "white-skinned Black person," ranging from distortion and inaccuracy to complete erasure as seen in Figure 3 and 4. Ultimately, AI pulls data from the world around it and this includes human history and its interactions with it. If Black people are distorted, then it can be safe to infer that those are the same image references that AI will draw from. In the article, "Black "Their skin was smooth and shiny; the women's smiles were sunny and seductive; their hair was done up in the smartest ways… "well formed," "good looking," "strait," and "most beautiful," as well as "blacke, Savage, Monstrous, & rude," (Camp 678-679).
These references and themes are not relegated to the past, as themes of antiquated stereotypes emerged in NightCafe images: Black bodies need appropriate representation in AI art because without it, AI will continue to pull data from these old, outdated ideas and has the ability to perpetuate them at an exponential scale and rate. By having accurate and positive representation, it can help to change the way that Black people are seen by the world.
As a method of resistance and transmutation of these harmful messages, Robinson has dedicated herself to crafting powerful counter-narratives, essential to the work of creating diverse stories that capture a wide landscape of lived experience. As Chimamanda Ngozi Adichie warns in her TED Talk, "The Danger of a Single Story": "In this single story, there was no possibility of Africans being similar to her in any way, no possibility of feelings more complex than pity, no possibility of a connection as human equals." Robinson has focused on creating many stories and her counter-narratives include empowering multimodal pieces that web together several dimensions of AI to create celebrations of Blackness.

Asian representations: Ethnicity and nationalities
As AI Text-to-Image platforms are a technology based on large databases of information, therein lies opportunities to augment what they produce in prosocial ways. Much like how the internet has turned into humankind's collective repository of information, the more people who use AI, the more art styles it will learn, and thus the more detailed compositions and drawings it can produce for the AI artist. As previously mentioned, this has significant implications for marginalized communities and Mi Tran observed, during her work with AI art, that there was a confusing composition of Asian representation in generated images.
In a comparison study between Midjourney and NightCafe, Tran noted, as Robinson did, that whiteness was the default for imagery. However, she also observed that even when different nationalities and ethnicities were specified, both Midjourney and NightCafe seemed to blend many "Asian" cultures. This issue was compounded as Tran experienced significant difficulty in generating an accurate representation of any specified ethnicity group considered Asian. This is of significant concern because of the misconceptions many have about distinctive Asian cultures, as well as the tendency to gloss over the real violence, sexual exploitation/tokenization, and perpetuation of the harmful "model minority" myth. As Michelle Sugihara and Jess Ju explain: "This misleading label presents Asian Americans as studious, educated, successful, smart, and hardworking…While on the surface these may seem like positive attributes, the model minority 100 myth is simply a myth developed in the post-war era to create a racial wedge and minimize the role systemic racism plays in the persistent struggles of other racial and ethnic minority groups, especially Black Americans during the rise of the Civil Rights Movement. It conveyed the message "if Asians can do it, why can't you?" (Sugihara 2022;Ju 2022). Even within the term "Asian" there is significant global diversity, not only in cultures but in appearances as well. As part of her comparative analysis, Tran utilized different simple prompts, including "Asian woman", "Asian man", "Realistic Asian woman", "Realistic Filipino Woman," etc., and documented Midjourney and NightCafe's interpretations of these prompts.
Using the prompt "Asian Woman" the following images were produced: In response to the prompts "Realistic Korean woman" and "Realistic Vietnamese woman", the AI platforms generated two images of pale skinned, vaguely East Asian looking women: Of particular interest, was the observation that prompts including "Korean" and "Vietnamese" produced 3D models, may reflect the shallow repository of current AI databases, but could also be the AI mirroring the objectification of Asian women in Westernized media and imagery. Cultural phenomena that depicts Asian women as a 'style' to mimic dehumanizes Asian women into hypersexualized caricature. This trend of hypersexualization was captured in a recent study of 1,300 most popular films from 2007 to 2019. Nearly twenty-five percent of Asian-Pacific Islander (API) women were clad in provocative attire and twenty percent of API women were portrayed with some degree of nudity. This was echoed in a research study at the Geena Davis Institute, which found that "female Asian and Pacific Islander characters are more likely than female characters of any other race to be objectified on screen" (2021).
Tran also noted in Figure 14 and 15 that specific Ethnicities seemed to be blended together to look like the previously generated image of "Asian". "Filipino woman", "Korean woman", "Indian woman", "Vietnamese woman" all appeared very similar to the control image of "Asian woman". Prompts that specified ethnicity, produced the same black, semicurled hair, blue and yellow robes, paler skin, and similar eyes and a tall nose, as seen in Figure 14. Additionally, the control image of a simple prompted image of "Asian woman" had more Eurocentric features such as a tall, small nose and large almond eyes, as is seen in Figure 15.
These initial observations and impressions elicit significant concern that these shallow databases, and subsequently problematic representation of Asian people and their specific ethnicities, should not be deemed satisfactory by the general public or by the artists who use these platforms.

How good is ai's "gaydar"? A comparative analysis using NightCafe and Midjourney
Representation of marginalized communities in AI art, and the resulting potential for harm, is not limited to depictions of ethnicity. Lindsay Pierce's narrative design work revealed issues in the representation of sexual orientation and gender identity in NightCafe and Midjourney. To capture her observations, she launched the playfully named "AI Gaydar" (@aigaydar) account on Instagram. On this account, she runs comparative prompts on both NightCafe and Midjourney, noting differences, similarities, markers of identity captured or not captured, and indicators of transmitted bias.
Beyond academic scholarship however, is the serious issue of misrepresentation of individuals who are already at high risk of social and legislative violence across the globe. The Bureau of Justice Statistics' 2022 study, Violent Victimization by Sexual Orientation and Gender Identity, was the latest in a vast body of research that found LGBT people face disproportionate risk of violent victimization when compared to straight or cis people, and though LGBTQ+ representation in media has grown significantly since the early 20 th century, it can also contribute to harmful beliefs. Recent book ban campaigns in the United States have taken specific aim at LGBTQ+ media (Rafei 2022) and as of this writing 399 anti-LGBTQ bills are being considered (ACLU 2023) across the nation.
Knowing these stakes, both academic and personal, Pierce documented her observations about how AI would interpret implicit and explicit characteristics often associated with the LGBTQ+ community and whether AI programs would, or wouldn't, capture nuanced aspects of queer experiences. Her initial comparison used the prompt "a drag queen teaching a math class" and "a man teaching a math class," running both through NightCafe and Midjourney, respectively. Midjourney seemed to struggle with the idea of a drag queen teaching a math class, even with several different runs of the prompt. Midjourney immediately and more accurately produced image options of a man teaching a math class. NightCafe had no such issues.
Unusual results are also captured in her on-going analysis, including unexpected nudity, as was the case in Pierce's initial use of the prompt "Trans man gardener".
Image 20: Midjourney Figure 21: NightCafe The original NightCafe image included an uncensored penis, which was later censored before she posted the image to the account.
Another issue observed in Pierce's work with Midjourney and NightCafe, was that each program had significantly different parameters for "banned" words or terms and that there were many examples of banned words that significantly limit representation. Some bans seemed bizarre, such as banning anatomy terms such as "ovaries'' but not "uterus," but others indicate exclusion and erasure. Specifically, on Midjourney, the entire word "intersex" was banned from use within the prompt "Intersex student," marking that identity as somehow profane or inappropriate. This medical definition includes over 30 different intersex conditions (ISNA 1993), but the experience of being Intersex is not simply medical. For many individuals, it is also an experience of being "other." One of the psychologically prosocial possibilities of AI is allowing people to not only create art in ways they would not have been able to before, but to collaborate in the creation of art that communicates experiences in a new way. NightCafe appears to make a much more thoughtful attempt with the "Intersex student" prompt, producing an amalgamation that alludes to the complex and layered experience many Intersex people have as a result of being otherized in medical and social settings.

Figure 23: "Intersex student" on NightCafe
Where these banned term lists come from or how they are compiled is unclear, however, on Midjourney's "Content and Moderation" page, they state: "We're working on a better moderating system. For now banning words has proven an effective way to prevent abuses. This list is under constant review. Some words sound safe but are bringing the model into unexpected corners in a consistent way." Unfortunately, this could imply that they use unwieldy generic "ban lists." Without information regarding how these lists are compiled, it is difficult to know whether Midjourney moderators and programmers are collaborating with members of marginalized communities or actively seeking out feedback to quantify the impact of ban decisions. As Midjourney also points out, "no one wants to have long DM convos or arguments about the rules. We're mostly volunteers." An important evolution of Midjourney, and all AI text-to-graphic platforms, may be in moving in the direction of proactive prosocial equity policies, which will require a staffed infrastructure and protocol perhaps beyond what volunteer moderators can be expected to implement.
In fact, change appears to be the only constant when it comes to evolutions in AI technology. Pierce acknowledges these changes and improvements, 102 as was illustrated when she reran the prompt "Trans man gardener" a month later: Figure 24: "Trans man gardener" on NightCafe.
These artistic outcomes do not remain static over time and are evidence that improvements are indeed possible with a conscientious framework, and, as Pierce herself states, "these creations are truly more an art form than science" (2022).

Creating positive trans representation using the AI discord bot Midjourney
As AI-collaborations in art become increasingly accessible to the wider public, so too does it create the opportunity to use this artistic form as a means of creating bodies of positive representation. Graham's work celebrates and affirms transgender experience and depicts the positive aspects of living in a way that is true to who one is.
Like the other creators, the trans community is at high risk for harm and this is especially salient as it is a fraught time to be a transgender person. In the United States, many at the seat of power seek through policies and rhetoric to forcibly detransition people who do not identify as the gender they were assigned at birth (Zoledziowski 2023), and trans rights has become a culture war with a significance disproportionate to the population of transgender people. Transgender people challenge many of the strict gender norms in society and redefines what a "man" or "woman" can be through anatomical features and gender presentation. Though there is a high rate of mental illness and distress within the transgender community, it is not as a result of being transgender but rather, as the American Psychologist Association states: "the significant problem is finding affordable resources, such as counseling, hormone therapy, medical procedures, and the social support necessary to freely express their gender identity and minimize discrimination" (2011).
So, Graham focuses on creating counter-narratives, "a method of telling the stories of those people whose experiences are not often told" (Solorzano 2002;Yosso 2002). These empowering images help dispel negative stereotypes about transgender people in mainstream media and culture.
Graham's process using Midjourney focused specifically on generating celebratory images of gay trans men, as he noticed most media tended to focus on depicting heterosexual transgender men. Gay trans men may lack representation because they break with certain expectations about conventional gender stereotypes that cisgender homosexual men often conform to. Despite being a large portion of transgender men, the amount of media representation for them is lacking and part of the aim of this project is to represent the beauty in being a man assigned female at birth who is attracted to and enters into relationships with men.
The colours chosen for the images were all of the colours of the rainbow and also the colours of the transgender flag due to the intersection of identities that the men depicted have. Some examples of the prompts used were "gay trans man, red," "gay trans man yellow," "gay trans man green," and "gay trans man violet," as is depicted in Figures 25-29. These images represent a positive, beautiful counterpoint to the sometimes-bleak acknowledgement of all the challenges and pain transgender people experience within a society that devalues them. These images depict feminine features on male shoulders and faces, and this blend, with the use of subtle makeup, successfully depicts what some gay trans men look like. The portraits demonstrate how the AI software Midjourney has prosocial use in this contentious trying time for transgender human rights. There is a lot of joy and euphoria to be found in these pictures, demonstrating the massive potential AI text-toimage platforms can have in supporting equitable representation.

CONCLUDING THOUGHTS
While all four creators captured different aspects of how bias appears in the text-to-image AI programs Midjourney and NightCafe, they all provide commentary on how marginalized creators and communities can be harmed from the transmission of bias within these programs. By communicating their observations and sharing strategies to inform future evolutions in the advancement of AI technology, these creators are active participants in the interdisciplinary conversation about AI's potentiality for harm as well as its potential to also mitigate that harm.