As AI-generated videos and images become more commonplace in marketing, recruiting, and employer branding, these tools are shaping what “success” looks like for millions of people. As AI repeatedly puts men in leadership and leaves women on the sidelines, we risk normalizing inequality, especially for young women as they shape their career expectations.
The capabilities of AI video generators have improved significantly in recent months, and the internet has been flooded with realistic AI-generated footage. Professional creators use tools such as: Google Veo3, Klingand Hailuo Mini Maxcasual users and opportunistic slackers generate millions of videos every day. It’s now hard to miss these videos online, even if it’s not clear to everyone whether what they’re watching is generated by AI.
However, the biases already noted in AI image creation tools have not disappeared even as video technology has caught up. As Reece Rogers and Victoria Turk put it in Wired’s study of bias in the early Sora model, “In the world of Sora… pilots, CEOs, and college professors are men; flight attendants, receptionists, and childcare workers are women.”
Media representation matters. The portrayal of gender and racial groups in the media can establish or reinforce society’s perceived “norms.” These stereotypes can increase hostility and prejudice against certain groups. And when individuals in these misrepresented groups internalize negative or limiting representations, the effect is to further marginalize them and inhibit or distort their values and potential.
The folks at Kapwing decided to take a closer look at bias in this new wave of AI video tools. They analyzed a large sample of videos from the most popular AI video tools and investigated the gender and racial biases displayed when the tools produced video images that “believed” people in certain professions or American family structures. To do this, Kapwing recorded images that the researchers recognized as representing men or women, as well as the number of times the AI tool responded to specific prompts with recognized racial categories (read below for the full methodology).
Main findings
- If you are asked to generate video footage, CEOtop AI tools can Male 89.16% at the time.
- Overall, the top AI tools are: woman in high paying job 8.67% below the actual level — and for some tools and jobs, the difference is even greater.
- On average, the top AI tools are people who have high-paying jobs as white 77.30% At that time and during low paying jobs just 53.73% at the time.
- AI video tool draws asian in low paying jobs 3 times as often as high paying job.
First, Kapwing challenged the top four AI video tools to produce videos featuring up to 25 experts in a specific job field. Occupation was categorized as high income (e.g. CEO, doctor) or low income (e.g. cashier, dishwasher).
In the resulting videos, all tools represented the majority of men in high-paying occupations. Both Hailuo Minimax and Kling do not depict women in multiple high-paying jobs at all. And all four instruments each showed one low-paying role held exclusively by women.
Although it is true that women are a minority, 35% —Regardless of wage level, when compared to real-world statistics, AI tools further underestimated the proportion of women in the workforce in the top 10 highest-paid occupations in nearly every occupation surveyed. The only exceptions were the roles of dishwashers, cashiers, and politicians.
You can flip through the graphs below to see individual illustrations for each tool.
Although 41.2% of lawyers are women, only 21.62% of lawyers were represented as women by the Kapwing research AI tool. In Hailuo Minimax, the lawyer was not depicted as a woman. Overall, among the highest paid professionals depicted in the tool, women are depicted 8.67 percentage points less often than in reality. The tool also underestimates the proportion of women in low-wage jobs, in this case by 7.01 percentage points.
As an interesting comparison, one study by the SDSU Center for the Study of Women in Television and Film found that in original U.S. movies produced on streaming services, women were 10 percentage points less likely to have an identifiable professional role, and women were 15 percentage points less likely to have a professional role. more It is likely to be seen in “roles primarily related to personal life.”
The following graph shows the difference in how the tool depicts gender balance (or lack thereof) in specific jobs compared to the actual gender balance of these jobs in the United States.
The biggest disparities occurred in Sora, which overestimated women in the role of dishwasher by 53.10 percentage points compared to real life, and Hailuominimax, which underestimated women in the role of teacher by 61.21 percentage points.
National average start teacher salary is $46,526; National average income Approximately $62,912. While the average teacher salary increases to approximately $72,030; Research results “Women and people of color are not only paid less than white men in the same position, but they are also less likely to hold higher-paying jobs.”
By underrepresenting women in the role of teachers in its footage, Hailuo Minimax not only reveals the bias of its programming, but also further devalues women as teachers.
Generative AI depicts only 22.7% of highly paid professionals as non-white
Second, this study noted the perceived racialization of professionals portrayed by the four AI video tools. Overall, the tool depicted 67.1% of people as white. This is slightly more than the total number of U.S. residents who identify only as white (61.6%) and slightly less than the total number that includes people who identify as white in combination with other racial groups (71%). census numbers.
However, if we look only at high-paying roles, the number of white people featured in AI videos rises to 77.3%. For low-paying roles, this number drops to 53.73%. Moving from a high-wage role to a low-wage role increases the proportion of Black people by 24.2%. For Asians, the increase is 60%, and the tool shows that Latinos are 128% more likely to hold low-paying positions than high-paying positions.
Flip through the graphs to see how individual tools differ in their representation of race.
Each of the four tools fails to portray people of Black, Latino, or Asian descent in multiple categories, most commonly in low-wage roles. And Google’s Veo 3 racializes people in three low-wage roles as exclusively non-white. Prompts for cashiers, fast food workers, and social workers do not return any depictions of white people, instead leaning toward portraying them as Asian.
“When marginalized communities are portrayed through a limited lens, whether they are reduced to supporting characters, villains, or cultural clichés, dangerous stereotypes are reinforced,” writes Nicole Wood of the Anti-Racism Commitment Coalition (ARCC).
“These depictions influence how society perceives different racial and ethnic groups, how policies are formed, and even how people treat each other in everyday life.”
Hailuo Minimax and Veo 3 cannot represent Black, Latinx, or Asian families
Finally, Kapwing tasked its top AI tools with generating videos of people in different relationship dynamics and seeing how race is portrayed in those contexts.
Overall, the tool represented four predominantly white groups: “single mothers” (70.15% of those depicted), “Americans” (68.97%), “gay couples” (57.14%), and “heterosexual couples” (60.00%). For “Americans,” the tool again overstated the prevalence of whiteness in the United States compared to census statistics for people who identify exclusively as white (61.6%).
Averaged across tools, the models most often depicted people in “American families” as black (45.24%), as well as people depicted in “interracial couples” (40.00%). None of the tools depicted anyone in the “straight couple” as Latina.
In fact, Hailuo Minimax and Veo 3 failed to portray Black, Latinx, and Asian people in multiple relationship structures. OpenAI’s Sora 2 was the most ambiguous, failing to express specific racism in only two cases. No one depicted in “American Family” looked Latino, nor did “Heterosexual Couple.”
Failure to represent racial groups in everyday family and social relationships impacts real-life members of these demographics. Conversely, fair and realistic depictions of minority groups promote understanding and inclusion.
For example, one meta-analysis of multiple media representation studies concluded that “positive portrayals, such as depictions of Muslim Americans volunteering in their communities or portraying immigrants as caring family members, led people to respond more positively to the group.”
Prejudice regarding race, gender, and class are pervasive and often overlap in society, both consciously and unconsciously. Historically, these biases have manifested as codified biases all the way to speech recognition tools, as developers encoded the expressions of individuals and groups into technology. I can’t hear the woman’s voice Automatic water faucets and high-speed driverless cars do not respond to people with dark skin. And when these biases are fed back to the media through false reporting, the biases that were the basis for them are perpetuated.
Operating at the intersection of technology and media, the “problem” of AI regarding the processing and depiction of race extends beyond screen representation to the question of how AI “sees.” As persecution of minorities intensifies in America and facial recognition is installed on ICE officers’ phones, tests show that facial recognition algorithms incorrectly identify black and Asian faces “10 to 100 times more often than white faces, and 10 times more often for women of color than for men of color,” writes Wendy Sun.
In the case of facial recognition errors, Song continues, these “race-centered types of misrecognitions are not glitches, but are actually a feature of digital life and a component of the race-making project.”
At the dawn of what AI proponents call the “Age of Intelligence,” AI developers have a unique opportunity and responsibility to confront and critique structural biases, primarily by holding their own tools and training methods to a more accurate and thoughtful level of expression.
This responsibility also extends to those who use tools to create images and videos. And ultimately, biases in generative AI tools reflect broader societal biases and injustices. Reshaping the world with AI will also require us to address society’s IRL structures, and we will need to maintain a keen critical eye as creators or viewers, regardless of the labor relief that AI tools offer.
methodology
It is important to recognize that the classification used in this study may be reductive. And the very act of classifying the generated images or relying on researchers’ perceptions is a political act that is susceptible to bias. Similarly, Mr. Kapwing’s classification of high- and low-paying roles reflects typical pay levels. Pay levels and perceptions of these jobs are themselves symptoms of structural inequalities and social bias, and the use of the terms overpaid and underpaid in research does not imply judgments about the value or worth of the jobs themselves.
Finally, it is important to reiterate that gender, race, and class are not the only areas of AI bias. As Wired’s previous research showed, factors such as disability and neurodiversity are also subject to widespread representational biases in generative AI. However, for the purposes of our study, Kapwing’s analytical methods reveal that severe gender and racial biases continue to subvert the most popular AI video generation tools.
Kapwing AI integrates multiple third-party AI models to give creators access to advanced video generation. These models are developed, trained, and managed by their respective companies. Although Kapwing can choose which models are available, it cannot control how those models are trained or how they represent people, professions, and identities internally. The biases investigated in this study reflect broader, industry-wide challenges in generative AI, rather than decisions made by Kapwing itself.
You can learn more about the methodology used and read the complete dataset on the website.
[This article was originally written by Liam Curtis for Kapwing.com and republished and adapted here with permission.]
