VISUAL SUPPLEMENT FOR BLONDE BRAIDS STUDY (2022) BY MINNE ATAIRU
This gallery is a visual supplement for my AI-generated portrait
series titled Blonde Braids Study (2022). The series began with a question:
Can the text-to-image algorithm — Midjourney (V4) — accurately generate studio portraits of "blue-black"
or
"plum-black" complexioned
twins
sporting blonde braids?
Contrary to the envisioned
prompt, the algorithm generated portraits of fraternal twins with
rolled, permed and silky hairstyles. Their complexions recurred in two variations: [ 1 ]
a juxtaposition of caramel and chocolate hues, and [ 2 ] matching caramel complexions.
Given these outputs, one might question: Why does Midjourney (V4), despite its advanced capabilities,
diverge from the given description?
Bender et al. (2021) share critical insights: Generative systems (such as
Midjourney) are designed to synthesize images based upon patterns
discerned from large-scale, uncurated datasets. Amassed through indiscriminate web scraping, such datasets
are shaped by the dynamics of internet participation, and therefore, represent the viewpoints of those with
consistent internet access. "In the
case of US and UK, this means that white supremacist and
misogynistic, ageist, etc. views are overrepresented in the training
data, not only exceeding their prevalence in the general population
but also setting up models trained on these datasets to further
amplify biases and harms."
This insight prompts further
interrogation: Do the above-described portraits
mirror any biased perspectives?
About
Black hairstyles? To what extent are such perspectives symptomatic of natural hair
discrimination - a form of bias marked by the
hyperregulation and hypersurveillance of Black hair textures?
From the era of chattel enslavement to
present-day, systems of oppression have shaped societal attitudes towards natural Black hairstyles such as
afros, braids, twists, locks. Black hair is often per·ceived as 'bad' hair, 'woolly' hair, 'bushy' hair,
'unprofessional' hair, 'dreadful' hair, as indicative of the “angry Black woman" stereotype, as a marker of
criminality. Consequently, Black folks feel pressured (Dawson et al., 2019) to conform to Eurocentric
hairstyles by straightening/smoothing their naturally curly/coily hair. Yet, when they choose
blonde-colored hairstyles — whether achieved through protective styling, chemical/thermal treatments, or
synthetic extensions — this choice is deemed [ 1 ] "unnatural" for their racial identity, and [ 2 ] a
violation of "hair privileges" traditionally reserved for Whiteness (Greene, 2010).
Such biases are evident in arbitrary workplace grooming policies. For example, in 1999, Andrea Santee-a "Black woman with dyed
blonde
hair" faced
employment
discrimination at Windsor Court Hotel in New Orleans. Santee's blonde hair
was categorized as "an extreme hair color" under the hotel's grooming policy.
In 2003, Shirley
Bryant—a CUNY (City University of New York) employee with
short curly blonde hair, experienced workplace harassment for deviating from "Afrocentric" hairstyles.
Bryant's supervisor allegedly called her a "wannabe" (meaning: "want to be White"). In
2009, Dulazia
Burchette—an employee at Abercrombie and Fitch was pressured to remove blonde highlights or resign
(Greene, 2010). Burchette's supervisor allegedly said: "she should have the hair color she was born with". In
2017, Destiny
Tompkins—a Banana Republic employee was subjected to workplace discrimination for sporting "urban" and
"unkempt"
blonde box braids. In 2019, Marion
Scott—a student at Paragon Charter Academy was excluded from
school photos due
to red-colored braids (Abrahamson, 2019).
Building upon these observations, I delved into a secondary inquiry:
Does Midjourney's training data (LAION-5B) include images of melanin-rich folks wearing
blonde
braided hairstyles?
To explore, I conducted a NSFW-filtered
image search on LAION’s
database for two queries—'Blonde Box Braids' and 'Blonde Braids' (dated: November 11, 2023). The
results overrepresented images of caramel-complexioned Black women and White women. "Blue-black",
"plum-black", "chocolate","dark-skinned", "melanin-rich" women were underrepresented in the homogenized
dataset. The search
results — a collection of over 10,000 images — are viewable on this site.
TECHNICAL NOTES
LAION-5B is a dataset of over 5 billion images and their corresponding captions. The dataset is
searchable through a portal
(now-defunct),
which leverages the capabilities of CLIP (Contrastive
Language-Image
Pre-training) for image retrieval. Unlike conventional search
engines that
rely on exact keyword matches
, CLIP assesses images based on their conceptual relevance to text queries.
This approach allows for the retrieval of images based on CLIP's interpretation of the
terms rather
than exact keyword matches. As such, the image results for the search queries—'blonde braids', and 'blonde
box
braids' include captions that do not explicitly mention these
terms.
None of the images displayed in this gallery are hosted on the site. They are
retrieved from external image URLs provided by LAION-5B dataset. All
inactive image links are
automatically
identified and replaced with lime green placeholders.
REFERENCES
- Abrahamson, R. A. (2019, October 8). See what happened after a girl was
denied a class picture because of her hair. Today. Retrieved from
https://www.today.com/parents/michigan-girl-denied-yearbook-photobecause-her-hair-t164152
- Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the dangers of stochastic
parrots: Can language models be too big?🦜. In Proceedings of the 2021 ACM conference on fairness,
accountability, and transparency (pp. 610-623).
- Greene, D. W. (2010). Black women can't have blonde hair in the workplace. J. Gender Race & Just., 14,
405.
- Dawson, G. A., Karl, K. A., & Peluchette, J. V. (2019). Hair matters: Toward understanding natural black
hair bias in the workplace. Journal of Leadership & Organizational Studies, 26(3), 389-401.
- MacLin, M. K., & Herrera, V. (2006). The criminal stereotype. North American Journal of Psychology, 8(2),
197-208.
- Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., ... & Jitsev, J. (2022).
Laion-5b: An open large-scale dataset for training next generation image-text models. Ad