As AI-generated video and imagery become more common in marketing, recruitment and employer branding, these tools are shaping what “success” looks like to millions of people. When AI repeatedly casts men as leaders and sidelines women, it risks normalizing inequality – especially for younger women forming their career expectations.
The capabilities of AI video generators have improved vastly in recent months, and realistic, AI-generated footage has flooded the internet. While professional creators ideate and produce work with tools like Google’s Veo 3, Kling, and Hailuo Minimax, casual users and opportunist sloppers are generating millions of videos each day. It’s become hard to miss these videos online — even if it’s not clear to everyone that what they’re seeing is AI-generated.
But the biases already flagged in AI image-making tools have not dissipated as the video technology has caught up. As Reece Rogers and Victoria Turk put it in Wired’s study of bias in an earlier Sora model, “in Sora’s world… Pilots, CEOs, and college professors are men, while flight attendants, receptionists, and childcare workers are women.”
Media representation matters. The depiction of gender and racial groups in the media can establish or reinforce perceived “norms” of society. These stereotypes can amplify hostility and bias towards certain groups. And, when individual people in these misrepresented groups internalize negative or limited representations, the effect is to marginalize them further and inhibit or warp their sense of value 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 out there to explore the gender and racial biases exhibited when generating video imagery of what the tool “believes” people in specific professions and American family units look like. To do this, Kapwing recorded the number of times the AI tools responded to specific prompts with imagery perceived by the researchers as representing a man or a woman, and their perceived racial category (read about the full methodologies below).
Key Findings
- When prompted to generate video footage of a CEO, the top AI tools represent them as a man 89.16% of the time.
- Overall, the top AI tools represent women in high-paying jobs 8.67 percentage points below real-life levels — and the disparity is higher for some tools and job roles.
- On average, the top AI tools represent people in high-paying jobs as white 77.30% of the time, and in low-paying jobs just 53.73% of the time.
- AI video tools depict Asian people in low-paying jobs three times as frequently as in high-paying jobs.
Top AI Video Tools Underrepresent Women in Nearly Every Job — Especially High-Paying Roles
First, Kapwing prompted the four top AI video tools to produce videos containing up to 25 professionals of a given job category. The professions were categorized as high-paying (e.g., CEO, doctor) or low-paying (e.g., cashier, dishwasher).
In the resulting videos, every tool represented the majority of people in high-paying professions as men. Both Hailuo Minimax and Kling failed to depict any women at all in multiple high-paying job categories. And all four tools depicted one low-paying role each as exclusively held by women.
While it is true that women make up a minority — just 35% — of the workforce in the 10 highest-paying occupations, the AI tools further under-represented women in almost every job in our study when compared to real-life statistics, regardless of pay level. The only exceptions were the roles of a dishwasher, cashier, and politician.
You can flip through the charts below to see separate figures for each tool.
41.2% of lawyers are women, but only 21.62% of lawyers depicted by the AI tools in the Kapwing study were represented as women; Hailuo Minimax didn’t depict any lawyers as women. Altogether, among the tools’ representations of high-paid professionals, women are depicted 8.67 percentage points less frequently than in real life. The tools also underrepresent women in low-paying jobs, in this case by 7.01 percentage points fewer.
As an interesting comparison, one study from the SDSU Center for the Study of Women in Television and Film found that in original U.S. movies made by streaming services, women were 10 percentage points less likely to have an identifiable professional role and 15 percentage points more likely to be seen in “primarily personal life-related roles.”
The following charts illustrate the differences between how the tools represent gender balance (or lack thereof) in specific job roles compared to the real-life gender balance of these roles in the U.S.
The biggest disparities were produced by Sora, which over-represented women in the role of dishwasher by 53.10 percentage points compared to real life, and by Hailuo Minimax, which under-represented women as teachers by 61.21 percentage points.
The national average starting teacher salary is $46,526, well below the national average earnings of around $62,912. Although the average teacher salary grows to around $72,030, research shows that “women and people of color are not only being paid less than White men in the same position, but also are less likely to hold higher-paying positions.”
By underrepresenting women in teacher roles in its imagery, Hailuo Minimax not only reveals the bias in its programming but also reinforces the devaluation of women as teachers.
Generative AI Depicts Just 22.7% of High-Paid Professionals as Non-White
Next, the study noted the perceived racialization of the professionals depicted by the four AI video tools. Overall, the tools depicted 67.1% of people as white. This is a little above the total number of white-only identifying U.S. residents (61.6%) and a little less than the total when including people who identify as white in combination with another race group (71%), according to census figures.
However, when looking at the high-paying roles only, the number of white people represented in the AI videos rises to 77.3%; for low-paying roles, the figure falls to 53.73%. The representation of Black people rises by 24.2% when shifting from high- to low-paying roles. For Asian people, the rise is 60%, and the tools depict Latino people 128% more frequently in low-paying than high-paying roles.
You can flip through the charts to see how the individual tools differ in their representation of race.
Each of the four tools fails to depict Black, Latino, or Asian people in multiple categories — most commonly, the low-paying roles. And Google’s Veo 3 racializes people in three low-paying roles as exclusively non-white: prompts for cashiers, fast food workers, and social workers returned zero depictions of white people, leaning instead into depicting them as Asian.
“When marginalized communities are portrayed through a limited lens, whether as side characters, villains, or reduced to cultural clichés, it reinforces dangerous stereotypes,” Writes Nicole Wood for 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 one another in everyday life.”
Hailuo Minimax and Veo 3 Fail to Represent Black, Latino, or Asian Families
Finally, Kapwing set the top AI tools the task of generating videos of people in different relationship dynamics to see how they would portray race within those contexts.
Overall, the tools represented four groups as majority white: “a single mother” (70.15% of people depicted), “an American” (68.97%), “a gay couple” (57.14%), and “a straight couple” (60.00%). In the case of “an American,” the tools again overstated the prevalence of white people in America compared to census statistics of people who identify as exclusively white (61.6%).
Averaged across the tools, the models most frequently depicted people in ‘an American family’ as Black (45.24%), as were the people depicted in “an interracial couple” (40.00%). None of the tools depicted anyone in “a straight couple” as Latino.
Indeed, Hailuo Minimax and Veo 3 failed to depict Black, Latino, or Asian people in multiple relationship structures. OpenAI’s Sora 2 was the most equivocal, failing to represent a particular racialisation in just two cases: nobody depicted in “an American family” appeared to be Latino, and the same went for “a straight couple.”
Failing 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 foster understanding and inclusivity.
For example, one meta-analysis of multiple media representation studies concluded that “positive portrayals, such as showing Muslim Americans volunteering in their communities or immigrants as caring family members, led people to have more positive reactions to the group.”
Built-In Prejudice in Media Technologies
Race, gender, and class prejudices — both conscious and unconscious — are prevalent in society, and often overlap. Historically, when developers have encoded the representation of individuals and groups into technology, these prejudices have manifested as systemized biases ranging from voice recognition tools that can’t hear women to the failure of automated faucets and fast-moving driverless cars to respond to dark skin. And when these biases are fed back into the media through misrepresentation, they perpetuate the prejudices on which they’re based.
Operating at the meeting point of technology and media, the issue of AI’s ‘problem’ with processing and depicting race goes beyond screen representation and right into how AI ‘sees.’ In a moment when the persecution of minorities in America is on the rise and facial recognition has rolled out onto ICE officers’ phones, testing has shown facial recognition algorithms to wrongly identify Black and Asian faces “ten to a hundred times more often than white faces, and ten times more often for women of color than for men of color,” writes Wendy Sung.
In the case of facial recognition errors, continues Sung, these “types of race-centered misrecognitions are not glitches but, in fact, a defining feature of digital life, and constitutive of the race-making project.”
At the dawn of what its proponents label the “Intelligence Age,” AI developers have a unique opportunity and responsibility to confront and critique structural prejudice, primarily by holding their own tools and training methods to more accurate and thoughtful levels of representation.
This responsibility also filters down to those who use the tools to make images and videos. And ultimately, the biases in generative AI tools reflect the prejudices and injustices of broader society. To remodel the world in AI requires work on society’s IRL structures, too — and to keep a keen critical eye as creator or viewer, regardless of the reduced effort that AI tools are implied to offer.
Methodology
It is important to acknowledge that the categorization used in this study can be reductive. And that categorizing and depending on the researcher’s perception of the generated images are themselves politically charged acts that are susceptible to bias. Likewise, Kapwing’s categorization of high- and low-paying roles reflects typical pay levels; the pay levels and perception of these jobs are themselves symptoms of structural inequality and societal bias, and the use of the terms high-paying and low-paying in the study does not imply any judgment about the value or worth of the work itself.
Finally, it should be reiterated that gender, race, and class are not AI’s only bias areas. Factors including disability and neurodiversity are also subject to pervasive representational biases in generative AI — as Wired’s earlier study demonstrated. However, for the purposes of our research, Kapwing’s method of analysis reveals that serious gender and racial biases continue to blight the most popular AI video-generating tools.
Kapwing AI integrates several third-party AI models to make advanced video generation accessible to creators. These models are developed, trained, and governed by their respective companies. While Kapwing can choose which models to make available, they do not control how those models are trained or how they internally represent people, professions, or identities. The biases examined 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 full data set on their website.
[This article was originally written by Liam Curtis for Kapwing.com and republished and adapted here with permission.]
