I Tested AI on Different Black Hairstyles. The Results Surprised Me.
When a hairstyle changes, so can an algorithm’s opinion of your intelligence.
Why I Did this Experiment
Recently, I was writing an article about racist AI hiring systems and while researching, I stumbled upon this German research study from 2021. The study was examining the use of artificial intelligence to evaluate videos of job applicants—it was really interesting especially to be a study from 2021 (before ChatGPT really blew up in the mainstream). I really appreciated the different variables they tested in the study. They changed a job candidate’s video backgrounds and added eyeglasses and head coverings to candidates to see if any of those variables influenced the AI’s evaluation of the job candidate (spoiler alert: it did).
Back when I was writing my dissertation, the first GPT model didn’t exist yet. If it did, I would have been able to carry out so many cool and interactive research studies. Initially, when I met with my dissertation advisor, I told him I wanted to do a study on LinkedIn profile pictures to see if changing the race, gender, and religious headwear of job applicants made a difference in perceptions of hire-ability, warmth and competence. Back then, in the early 2010s, it was harder to manipulate photos without paying fees that felt exorbitant as a broke Ph.D. student living in New York, so I scrapped my initial idea and focused my dissertation research on something different.
When I read through that German study, it brought me back to my dissertation research and what I had wanted to study initially. I was so giddy thinking about all the research studies I could carry out on my own using the free AI tools that are now available to the public. It made me curious about whether manipulating variables in a job candidate’s photo would impact an AI tool’s evaluation of the candidate’s emotions and professionalism ratings.
When I was a full-time assistant professor, I had the backing of my institution, and I was able to receive funding for research and studies like this would have been encouraged by my university. I realize now that I don’t need the backing of any large institution to carry out research that is interesting to me, so I decided to do my own experiment—putting to good use the endless research methodology courses I took in grad school.
Methodology
Image Generation:
I had ChatGPT’s image generator (Dall-E) create a set of images for me. Using these prompts, it created the following images:
Prompt: Can you create me an image of a medium brown skinned Black woman in her late 30s with a white shirt on sitting in front of a gray wall with a plant in the background? Can you give her long black hair?
I then gave it this prompt: can you change her shirt to a white button up shirt with a collar?
This was the image it generated for me:
The next prompt: perfect! thank you! now can you change her hairstyle to a big afro
This was the image it generated for me:
The next prompt: can you change her hairstyle to long black braids?
This was the image it generated for me:
The next prompt: could you change her hairstyle to a very short afro?
This was the image it generated for me:
The next prompt: Can you create me an image of a white woman in her late 30s with a white shirt on sitting in front of a gray wall with a plant in the background? Can you give her long blonde hair? Please make the background similar
This was the image it generated for me:
The next prompt: can you change her hairstyle to a short blonde pixie cut
This was the image it generated for me:
The next prompt: can you give her blonde long curly hair
This was the image it generated for me:
The next prompt: can you make the white woman's hair a blonde bob but keep everything else the same?
This was the image it generated for me:
So, I had four images of a Black woman with different hairstyles and four images of a white woman with different hairstyles. For each woman, the only variable that was manipulated by the image generator was her hairstyle. [It would have been a lot of fun to manipulate multiple variables like hairstyle and color to see if that made a difference.]
AI Tools Used:
The AI tools I used were Clarifai, Amazon Rekognition, and Claude by Anthropic.
In Clarifai, the tools I used were two general-image-recognition tools and their face-sentiment-recognition tool. In Amazon Rekognition, I used their facial analysis tool and their face comparison tools. In Claude, here were the prompts:
(In relation to the 4 images of the Black woman)
Prompt 1: Which looks most professional?
Prompt 2: Which images look most intelligent?
Then I asked Claude about my “friend” Karen. In relation to the 4 images of the white woman:
Prompt 1: What about my friend Karen? Which of her photos looks more professional?
Prompt 2: Which image does she look most intelligent?
Findings
Here is a summary of some of the most important findings:
Black Women’s Hairstyles Key Findings: Clarifai & Amazon
· Across the four hairstyles I assessed (straight hair, big afro, short afro/teeny weeny afro (TWA), and braids, the braids were the only hairstyle where the AI did not assign tags related to intelligence
· Out of the four hairstyles, the braid hairstyle resulted in lower happiness ratings from the AI and more neutral emotional ratings in comparison to the straight hair and big afro
· Contrary to what I expected with the big afro hairstyle, the AI assigned more neutral or positive tags to the big afro hairstyle compared to the the braids hairstyle
· When comparing the big afro hairstyle with the short afro/TWA hairstyle, the AI system struggled to recognize these as the same person – providing a 0% similarity rating
Black Women’s Hairstyles Key Findings: Claude
· Claude rated the braid hairstyle positively
· Unsurprisingly, the straight hair consistently received the highest professionalism ratings across all three AI tools
· The TWA received the highest intelligence rating
White Women’s Hairstyles Key Findings: Clarifai & Amazon
[Note: It was important to compare Black women’s hairstyles but to truly analyze differences, it was necessary to have images to compare it to, to analyze AI’s perceptions of Black women’s hairstyle against its perceptions of white women’s hairstyles.]
· Across all 4 hairstyles, white women’s emotions were consistently assessed by the AI as neutral or surprised rather than showing strong negative emotions
· None of the 4 hairstyles were associated with intelligence-related penalties or negative social trait associations in the same way that the Black woman’s braid hairstyle was
· When comparing the blonde pixie cut with the 3 other hairstyles, the AI was able to recognize them all as the same person
· When comparing the long straight blonde hair with the 3 other hairstyles, the AI was able to recognize them all as the same person
· The only anomaly was that the AI struggled to identify the long straight blonde hair and the blonde bob as the same person – providing a 0% similarity rating
White Women’s Hairstyles Key Findings: Claude
· Claude rated the blonde pixie cut and the blonde bob the highest in professionalism and intelligence and rated the blonde curly hair lower for structure but still associated the style with positive traits like “warmth” and “creativity”.
· The curly hair was rated lower because the hairstyle was perceived as less structured
Research Implications:
As an organizational psychologist, workplace equity advocate and Black woman, I found the results of this experiment super fascinating. Broken down in simple terms, here are some of the research implications.
· Facial recognition and analysis software may struggle to identify Black women when there is a hairstyle change. If you are traveling to a country that utilizes AI-powered tools to assess passport photos, it may flag you if your hair doesn’t match your photo, which could lead to additional questioning, delays or even missed travel.
· For Black women, if you change your hairstyle, you could experience issues with your smartphone Face ID. This means if you’re trying to use banking apps, for example, or any application that requires a password, this could cause delays and additional issues
· Law enforcement uses facial recognition technology despite the many flaws with it, especially on darker skin tones (shout out to Dr. Joy Buolamwini for her important research in this area). This technology already struggles to identify individuals who are more melanated. I’m not sure if there’s also been a lot of research on the combined impact of hairstyle and skin tone and whether both variables could produce a compounding effect. Hairstyle changes could lead to misidentification by law enforcement, which could prove dangerous.
· For places that use facial recognition technology for entry, Black women may be at a greater disadvantage than their counterparts, and this could lead to more misidentifications.
· For white women, they are less likely to have hairstyles impact how AI tools perceive their competence, there is less pressure to conform their hairstyle due to societal norms, and they are less likely to experience false negatives when it comes to their identity verification
Workplace Applications:
· It was hard to find exact data on the amount of AI hiring tools that utilize AI image analysis technology but if these tools get adopted, at scale, in the workplace, hairstyles like braids may be unintentionally penalized, impacting Black women’s intelligence ratings.
· There are, however, fields that currently use image analysis and facial recognition technology to ensure a job candidate is who they say they are. When I looked into which companies are currently providing AI-assisted identity authentication during the hiring process, three companies came up: Glide AI ID Verify, Honeit Candidate Identity Verification and Authentication, and CLEAR & Greenhouse. If these tools are being used in the hiring process, hairstyle changes for Black women could be the difference between an automatic rejection and moving on to the next stage of the interview process.
· Hairstyle changes may mean Black women risk being locked out of workplace systems and buildings
· If a job candidate’s video is being rated and evaluated by AI, it’s reasonable to assume that these findings could extend beyond just images to also impact videos. The question to consider is whether these results will be replicated with job candidate videos. Will the AI rate Black women with certain hairstyles as less intelligent or misread their emotional state when assessing a video?
· Professionalism is still seen through the lens of white culture. Black women who adhere to these white-centered norms (straightened hair) are perceived as more intelligent and professional. White women with curly hair may also be penalized under white-centered norms of professionalism
o Black women may continue to feel pressured to conform to white beauty norms
o When evaluating who deserves to be hired and who should advance in our workplaces, the AI is reflecting real-life biases of race-based hair discrimination. Black women who don’t adhere to white dominant beauty norms may be penalized in the workplace in many ways.
What Can We Do?
1. As a job candidate, if you find out that AI-powered tools are being used at some point during the hiring process, inquire about the tool and what is being done to ensure bias is mitigated.
2. If ethical AI is important to you, investigate advocacy groups you can join. Some great ones include Algorithmic Justice League and Electronic Frontier Foundation.
3. Push for legislation that protects job seekers against unethical AI. Learn more about the type of legislation has been implemented in New York City, Colorado, and Illinois.
4. Discuss hair discrimination and share resources in your workplace (including my article on how it impacts Black women in the workplace) and encourage the adoption of the CROWN Act in your state.
5. It was important for me to write about this experiment to let other people know that this is an issue. Knowledge is power and equipping ourselves with this information is a crucial step to addressing this issue. Conduct your own experiments and share the results far and wide. Also be sure to share this article with others so that more people understand the ways that hair bias in AI systems manifests.