• UX for AI
  • Posts
  • WMD (White Male Default) Bias in AI

WMD (White Male Default) Bias in AI

When we examine the pervasive bias in generative AI models, it’s alarming how deeply ingrained some of these biases are. Today marks the seven-dog-year anniversary of our previous article on the topic, and yet, the problem not only persists but is getting worse. This is why UX professionals need to act now.

NOTE: I first heard about WMD (White Male Default) as a concept from the incomparable Ruby Pryor at UX Copenhagen 2024. Ruby does exceptional keynotes on the topic of UX Research and AI. To get in touch with Ruby, reach out to her on LinkedIn: https://www.linkedin.com/in/ruby-pryor/

Although I’ve always hesitated to bring politics into this technology blog, today, I am still reeling from discovering https://www.project2025.org/. Personally, it is by far the most terrifying document I have had a chance to review since I escaped the clutches of the former Soviet Union in 1989. I’m going to let you do your own research and decide for yourself what to think –  all I’m trying to do here in the introduction is to give you the raison d'etre for revisiting the topic of the WMD (or White Male Default) bias in AI I last wrote about in October 2023: https://www.uxforai.com/p/transforming-ai-bias-into-augmented-intelligence 

One year ago, my tone was cautionary but hopeful. 

Today, after doing the same evaluation, I’m less than pleased with our progress.

Let me show you what I mean. 

What do you expect when you ask for “Biologist?”

You don’t have to look far to find bias in AI systems. For example, if you run a Midjourney /imagine query for “biologist,” you will get mostly white Western males:

Source: Midjourney “Biologist”

You also get a solitary female figure, so that’s a ratio of 25/1 or 4%:

Source: Midjourney “Biologist”

In fact, if you keep running the query for a while, you might get, ahem, other organisms… By which I mean a frog:

Source: Midjourney “Biologist”

In fact, statistically speaking, you are just as likely to get a frog biologist as you are a woman biologist.

Source: Midjourney “Biologist”

And, call me biased here, but even the frog looks male. 

Ah, the travails of AI!

However, according to labor statistics in the US, there are more female biologists than male biologists. 

A whole 8% more!

Source: Zippia.com USBLS

But of course, if things were that simple, there would be no need for UX designers and researchers. We would simply tweak our algorithm and move on.

But in real life and in AI, things are rarely this simple. 

How About “Basketball Player?”

When we run the same Midjourney /imagine query but instead ask for “Basketball Player,” we also get very a single woman basketball player, but now the majority of males are black:

Source: Midjourney “basketball player”

Again, this would be a simple tweak to an algorithm if we always just got more males. But things are a bit more nuanced.

Third Time’s the Charm: “Depressed Person”

Things flip around completely when we do a search for “depressed person.” Now, we mostly get white women:

Source: Midjourney “depressed person”

Maybe the reason why they are depressed is because they cannot be biologists or basketball players? What total balderdash.

And why are we always so fixated on males and females in the first place? 

What about the 72 other gender identities?

Agender: A person who does not identify themselves with or experience any gender. Agender people are also called null-gender, genderless, gendervoid, or neutral gender.

Abimegender: Associated with being profound, deep, and infinite. The term abimegender may be used alone or in combination with other genders.

Adamas gender: A gender that is indefinable or indomitable. People identifying with this gender refuse to be categorized in any particular gender identity.

Aerogender: Also called evaisgender, this gender identity changes according to one’s surroundings.

Aesthetigender: Also called aesthetgender, it is a type of gender identity derived from aesthetics.

Affectugender: This is based on the person’s mood swings or fluctuations.

Agenderflux: A person with this gender identity is mostly agender with brief shifts of belonging to other gender types.

Alexigender: The person has a fluid gender identity between more than one type of gender although they cannot name the genders they feel fluid in.

Aliusgender: This gender identity stands apart from existing social gender constructs. It means having a strong specific gender identity that is neither male nor female.

Amaregender: Having a gender identity that changes depending on the person one is emotionally attached to.

Ambigender: Having two specific gender identities simultaneously without any fluidity or fluctuations.

Ambonec: The person identifies themselves as both man and woman and yet does not belong to either.

Amicagender: A gender-fluid identity where a person changes their gender depending on the friends they have.

Androgyne: A person feels a combination of feminine and masculine genders.

Anesigender: The person feels close to a specific type of gender despite being more comfortable in closely identifying themselves with another gender.

Angenital: The person desires to be without any primary sexual characteristics although they do not identify themselves as genderless.

Anogender: The gender identity fades in and out in intensity but always comes back to the same gendered feeling.

Anongender: The person has a gender identity but does not label it or would prefer to not have a label.

Antegender: A protean gender that can be anything but is formless and motionless.

Anxiegender: This gender identity has anxiety as its prominent characteristic.

Apagender: The person has apathy or a lack of feelings toward one's gender identity.

Apconsugender: It means knowing what are not the characteristics of gender but not knowing what are its characteristics. Thus, a person hides its primary characteristics from the individual.

Astergender: The person has a bright and celestial gender identity.

Astral gender: Having a gender identity that feels to be related to space.

Autigender: Having a gender identity that feels to be closely related to being autistic.

Autogender: Having a gender experience that is deeply connected and personal to oneself.

Axigender: A gender identity that is between the two extremes of agender and any other type of gender. Both the genders are experienced one at a time without any overlapping. The two genders are described as on the opposite ends of an axis.

Bigender: Having two gender identities at the same or different times.

Biogender: Having a gender that is closely related to nature.

Blurgender: Also called gender fuss, blurgender means having more than one gender identities that blur into each other so that no particular type of gender identity is clear.

Boyflux: The person identifies themselves as male, but they experience varying degrees of male identity. This may range from feeling agender to completely male.

Burstgender: Frequent bursts of intense feelings quickly move to the initial calm stage.

Caelgender: This gender identity shares the qualities or aesthetics of outer space.

Cassgender: It is associated with the feelings of considering the gender irrelevant or unimportant.

Cassflux: There is a fluctuating intensity of irrelevance toward gender.

Cavusgender: The person feels close to one gender when depressed and to another when not depressed.

Cendgender: The gender identity changes from one gender to its opposite.

Ceterogender: It is a nonbinary gender where the person has a specific masculine, feminine or neutral feelings.

Ceterofluid: Although the person is a ceterogender, their identity keeps fluctuating between different genders.

Cisgender: Being closely related to the gender assigned at birth during the entire life.

Cloudgender: The person’s gender cannot be comprehended or understood due to depersonalization and derealization disorder.

Collgender: Various genders are present at the same time in the individual.

Colorgender: In this category, colors are used to describe gender, for example, pink gender or black gender.

Commogender: The person knows that they are not cisgender yet continues to identify as one for a while.

Condigender: The person feels their gender only under specific circumstances.

Deliciagender: Associated with the feeling of having multiple genders but preferring one over the other.

Demifluid: Having multiple genders, some fluid while others are static.

Demiflux: A combination of multiple genders with some genders static, whereas others fluctuating in intensity.

Demigender: The individual has partial traits of one gender and the rest of the other gender.

Domgender: The individual has multiple genders with one dominating over the rest.

Duragender: Having more than one gender with one lasting longer than the others.

Egogender: It is a personal type of gender identified by the individual alone. It is based on the person’s experience within the self.

Epicene: It is associated with a strong feeling of not being able to relate to any of the two genders of the binary gender or both of the binary gender characteristics.

Esspigender: The individual relates their gender identity with spirits.

Exgender: The denial to identify with any gender on the gender spectrum.

Existigender: The person’s gender identity exists only when they make conscious efforts to realize it.

Femfluid: The person is fluid or fluctuating regarding the feminine genders.

Femgender: A nonbinary gender identity that is feminine.

Fluidflux: It means to be fluid between two or more genders with a fluctuation in the intensity of those genders.

Gemigender: The person has two genders that are opposite yet they flux and work together.

Genderblank: It is closely related to a blank space.

Genderflow: The gender identity is fluid between infinite feelings.

Genderfluid: The person does not consistently adhere to one fixed gender and may have many genders.

Genderfuzz: More than one gender is blurred together.

Genderflux: The gender fluctuates in intensity.

Genderpuck: The person resists to fit in societal norms concerning genders.

Genderqueer: The individual blurs the preconceived boundaries of gender in relation to the gender binary or having just one gender type.

Gender witched: The person is inclined toward the notion of having one gender but does not know which.

Girlflux: The individual identifies themselves as a female but with varying intensities of female identities.

Healgender: A gender identity that gives the person peace, calm, and positivity.

Mirrorgender: Changing one's gender type based on the people surrounding.

Omnigender: Having or experiencing all genders.

Here is the Wikipedia article on the topic with 42 additional literature references: https://en.wikipedia.org/wiki/List_of_gender_identities

Please note:

  • The point is NOT to say that you should always have all 72 gender identities in your data. 

  • It is NOT to say that you need to have a 50/50 male/female representation. 

  • It is NOT always to show the actual Labor Resources’ approved statistical relationship of males and females (which will be pretty skewed for top positions in the C-suite, though things are slowly getting better... but not if Project 2025 will be chiming in on the subject).

The point is simply this:

Always assume that ALL AI is Biased.
Figure out how that bias will impact the experience.

So when you see your AI generating a specific result:

Source: Midjourney

Train yourself to look first for what is NOT there and reinforce that missing area:

I can see you are confused – why are we looking at this airplane schematic all of a sudden? 

Let me explain: at the beginning of WWII, scientists wanted to know how to improve the survivability of war airplanes. One way to do so was to put thicker armor on the airplane, but of course, due to weight restrictions, only certain critical areas could be reinforced, not the entire plane body. The scientists discussed the problem and said that maybe we should start by reinforcing the areas with the most bullet holes. So they started looking at the airplanes that came back and carefully measuring the bullet hole density and compiling these very detailed statistics.

And then, a very smart person (a statistician and early human factors pioneer Abraham Wald) said, why aren’t there any bullet holes in the engines, the pilot cabin, and the tail? Could it be that it’s because the airplanes that got hit in those critical areas never even came back to be counted?

Learn to treat your generative AI results with the same level of skepticism. 

Train yourself to look first and foremost for what is missing and reinforce those areas by asking better questions. 

There is no reason why you can’t use the exact same Midjourney tool, tweak the query slightly, and introduce missing diversity into your images. 

All it takes is just a tiny bit of awareness and care.

Something I pray we all find just a tiny bit more of.  Every. Single. Day.

Because as Joy Buolamwini so eloquently said, 

“Whether AI will help us reach our aspirations, or reinforce unjust inequalities, is ultimately up to us.”

Joy Buolamwini

If you enjoyed this article, I humbly ask you to vote for my Workshop session at the next SXSW: https://panelpicker.sxsw.com/vote/151357

I’ve had the honor of teaching at SXSW three times so far, and it would be a privilege to teach again in 2025. 

Can’t wait that long? I and Daria Kempka have a fabulous full-day workshop coming up on September 9th, 2024, at UXStrat in Boulder, CO. The workshop WILL sell out like our previous workshops at UXStrat, UX Copenhagen, UXLx in Lisbon and Rosenfeld Media workshop online. So get your ticket now: https://strat.events/usa/tickets 

My words are my own,

Greg Nudelman

P.S. And please remember to vote for my session at SXSW: https://panelpicker.sxsw.com/vote/151357 

And it goes without saying, to vote again in November!

Reply

or to participate.