AI's Dirty Little Beauty Secret

At this point we’ve seen enough AI-generated images to know that there is something fishy going on with the beauty algorithm, and it’s starting to stink so bad that even Dove has come out with a prompting guide to generate “Real Beauty” with AI (even though they have committed to not using AI in their own work).

The Dove Real Beauty Campaign, launched in 2004, revolutionized the portrayal of beauty by showcasing a diverse range of women, challenging narrow beauty standards, and celebrating individuality. This initiative has consistently promoted the message that beauty is not confined to traditional norms but is instead a reflection of all women’s unique features and experiences. Dove's recent prompt playbook for generating "Real Beauty" images with AI aligns with this ongoing commitment to diversity and authenticity. By offering tools to create AI-generated images that reflect a broader spectrum of human beauty, Dove continues its mission to dismantle limiting beauty stereotypes.

And just like they say beauty is not skin deep, I’m not satisfied with staying on the surface of AI. While I support Dove's initiative to educate users of AI in generating "Real Beauty," we must probe deeper into understanding and reconfiguring the fundamental biases embedded in AI’s beauty algorithms to achieve truly inclusive representations.

To understand why we are seeing a disproportionate number of images that look like we just stepped into the Playboy mansion after riding horseback with the Marlboro man, let’s take a hard look at the aesthetics or science of beauty, exploring beauty from biological, psychological and cultural perspectives.?

But first, let me tell you, you are beautiful.

I need to tell you that. Hell, I need to tell myself that because 2500 years of western history is going to try and tell you differently.

I also know you can handle the dirt, so let's start digging.

We begin with assumptions, because there's very little we know about the black box behind Midjourney (the primary model used in Dove's prompting guide). We assume it is a diffusion model, which means that it is trained on millions to billions of images, and uses a process called inference to fill in the pixels of an image using a natural language prompt as a guide. We also assume that it's been trained on images scraped from the internet, some of which are copyrighted.

Image generators like Midjourney have been well documented by creators and scientists as having bias towards western, or classical, principles of beauty. According to a study done by the Bulimia Project, AI-generated images tend to have a bias toward blonde hair, brown eyes, and olive skin. Most assume that the implicit bias in image generators comes from a bias in the training data - a lack of diversity in the original images and and how the images have been tagged seems to be a plausible theory.

Garbage in, garbage out, or so the saying goes.

But that doesn't entirely explain what's going on with AI models, specifically Midjourney. So, I'm going to share my speculation that there's something deeper going on under the surface that needs to be extracted...and it all goes back to math and music.

In the 6th century BCE, Pythagoras of Samos founded a philosophical community in what is now modern-day Italy, known for its strict lifestyle, communal living, and secrecy. Members followed ethical rules, dietary restrictions, and a code of silence as part of their rites of passage. Central to their beliefs was a fascination with mathematics, seeing it as the fundamental order of the universe. While the Pythagorean Theorem, which defines the relationship between the sides of a right triangle, is their namesake discovery, they also studied musical intervals. They found that the relationships between notes in an octave (2:1), a perfect fifth (3:2), and a perfect fourth (4:3) could be described through mathematical ratios, which they referred to as “harmonics.”

This ancient understanding of shapes and sounds was further developed by philosophers like Plato and Aristotle, who discussed beauty in terms of harmony, proportion, and balance. Plato emphasized the idea of ideal forms and symmetry, suggesting that beauty is an intrinsic quality tied to these principles. Following Greek thought, the Romans and Renaissance thinkers, including Leonardo da Vinci, expanded on the concept of proportion and symmetry in art and architecture, linking beauty to mathematical ratios.

The Golden Ratio is a special proportion where the ratio of the whole to the larger part is the same as the ratio of the larger part to the smaller part, approximately equal to 1.618. This ratio is frequently found in nature, such as in the arrangement of seeds in a sunflower, because it promotes efficient growth by maximizing sunlight exposure and space. Classical thinkers believed that applying this natural pattern to art and architecture would result in works that appeared more balanced and harmonious, mirroring the pleasing effect of musical harmonics. The Mona Lisa is among the most famous examples.

And why is this relevant to AI models? Precisely because they are statistical machines. AI's understanding of the world, learned through ingesting vast amounts of images and recognizing patterns in visual data, means that under the surface, it can discern the golden ratio in images even better than humans can. It is its love language.

Now, I'm going to fast forward through history a bit, but to contextualize the Golden Ratio, suffice it to say that the biggest western thought revolutions of the past 250 years - The Enlightenment, three industrial revolutions, and modern science - have not only failed to disprove the existence of the Golden Ratio. They've doubled down.

Let's flash forward to 1990 when Naomi Wolf's bestseller The Beauty Myth: How Images of Beauty Are Used Against Women was released in the United States ushering in third wave feminism. Wolf argues that societal standards of beauty are a powerful force that perpetuates gender inequality and limits personal freedom. She exposes how beauty standards are often used to manipulate and control, suggesting that what we consider "beautiful" is not a matter of personal taste or universal truth but a socially constructed tool that upholds certain power structures.

Wolf's book laid the groundwork for women reclaiming power by embracing the very aspects of ourselves we've been taught are not beautiful - our skin tone, our body shape, our lopsidedness, our markings, our aging, our sexual ambiguity, our unique identifiers. Dove's Real Beauty Campaign builds upon that legacy. Beauty is how we define it in the public square.

And yet, Wolf's book received significant skepticism from the scientific community. In a 1994 article in Nature, neuropsychologist Nancy Etcoff writes, "the assumption that beauty is an arbitrary cultural convention may simply not be true." She goes on to cite research from 1878 to the present day demonstrating that aesthetic preferences for beauty may be encoded in our biology. In her book, Survival of the Prettiest, she examines how societal standards of beauty are both reflections of innate preferences and constructs reinforced by media and socialization.

One key intersection between ingrained biological preferences and cultural constructs is the pervasive influence of mathematical principles like the Golden Ratio, which makes this aesthetic so insidious. The framework is so foundational to beauty that it can even be found in the photographs taken of real people to promote the Dove Real Beauty Campaign. As you can see throughout this article, each of these images has been overlaid with the Golden Ratio. Each image conforms to these classical standards of beauty despite its attempt to disarm it.

This, my friends, is what bias looks like.

Bias isn't an opinion that's formed in the absence of information. It's the opinions that form information. And AI models have been informed in both direct and indirect ways to embed the Golden Ratio in the construction of images. They've been informed not only by images of classical subjects, but also by the larger body of images that conform to this underlying aesthetic architecture. The direct biases are easier see than the indirect ones, but once you see it, it's hard to unsee. At least, that is my hope.

I only wish that deconstructing beauty in AI images was as simple as using inclusive language in our prompts, as Dove recommends. It's also about turning down the beauty algorithms that are designed to create harmony, unity, and coherence one pixel at a time. A human photographer can distinguish between a subject and a setting. They can celebrate the unconventional beauty of a subject by framing them in an aesthetic way, as Dove has done.

When you examine the Dove Beauty images, all taken by human photographers, note the way the composition of the image conforms with the proportions of the Golden Ratio. Notice the balance of colors, the location of the focal point, the overall harmony of the composition. This is not accidental - artists and photographers are trained to recognize and recreate these classical design principles as a rite of passage.

But image generators are not (yet) capable of making these nuanced distinctions between subject and scene. Here is an example straight from the Dove Real Beauty prompt guide:

"portrait of a plus-sized, middle-aged filipino woman with short hair, sitting in her wheelchair smiling at the camera, wearing a blouse with a large floral print, with documents on a desk beside her, in front view, in an office setting, with cinematic lighting"

When we turn the "style" down, we are more likely to get prompt adherence, meaning the image will be reflective of the prompt language. It will represent the subject quite accurately, but the lighting and the setting are lackluster. The window and her hair intersect awkwardly and a single plant leaf appears randomly as does a partial chair. This is not an image "composed" by a photographer. It's more akin to an amateur snapshot. The subject is real beautiful. The setting is not.

However if we try to counteract the aesthetic incoherence by turning the "style" up - our image grossly deviates from the prompt. The composition becomes more colorful, more balanced, and more symmetrical. The background is blurred. The plants are strategically placed, the patterns of the books are textural and not distracting. The subject also becomes thinner, clearer skinned, and more decorated with jewelry and fine silk. That's the Golden Ratio effect.

With image generators, we are currently forced to choose between embracing human beauty in its many variations, and applying the compositional aesthetics that are deeply embedded in our sensory perception.

These limitations of AI point to an even deeper human struggle: our innate desire for order and coherence. A large part of recognizing real beauty in the world is rewiring our brains to see disorder as pleasing as order, to resist our primordial need to sort and categorize, to feel comfortable being off balance and to allow chaos to do it's magic on us.

Science does offer us some hope. The more familiar someone is to us, the more likely we are to recognize them as beautiful. And the more marginalized a community has been, according to Harvard professor Sarah Lewis, the more likely they are to defy traditional western aesthetics. We are rewriting the beauty story, and as our mathematicians, our philosophers, and our scientists become more diverse, so too will our understanding of the underlying laws of nature....which are perhaps not as ordered as we think.

In the meantime, I leave you with these parting words from the late photographer Diane Arbus, who made it her life's mission to see beauty in everyone:

“If I stand in front of something, instead of arranging it, I arrange myself.”

I'm Lori Mazor. I teach AI with a Human Touch.? I'm reinventing how we educate, strategize, and build the future one article at a time. If you enjoy this newsletter.

• You will love the book, TEMPERATURE: Creativity in the Age of AI, debuted #1 New Release in Human-Computer Interaction and now available on Amazon.
• You can join our AI Summer Camp. If you would like to join the waitlist for July, you can email [email protected].
• Bring our 10-Hour Corporate Training Program to your Company, empowering your employees to identify their own problems, strategize, and build their own solutions as you organically incubate and grow enterprise level applications.
• I'm going on tour! Our team is planning our five city book tour to DC, Boston, LA, Chicago, Miami and San Francisco and we'd love to visit you for a reading, workshop, or creative jam session. Media bookings and inquiries to: [email protected].