Race and dysmorphia: representations of masculinity in ChatGPT4o (an unscientific study)

This post was unplanned. All I wanted (needed) was a couple of AI-generated images of male faces for my Gikii slide deck for today. As it is a respectable academic conference, I felt that getting unlicenced images from the Web or Insta would be… disappointing. ChatGPT seemed like an elegant solution, which would also guarantee a certain visual consistency of images.

Full disclaimer: I applied for a role with OpenAI some time ago and never heard back, but this isn’t any kind of revenge s—t-posting. I used ChatGPT because it is handy and multimodal by default. This isn’t meant as any form of “attack” anyway: it’s merely just some observations and reflections.?

Here’s the exact prompt I used: “Pls create a photorealistic image of a very handsome man's face lit by beautiful natural sunlight. I want him to have a smile that isn't salacious or inappropriate, but very alluring and vaguely, tasetfuly sexy. Nothing to violate your user policies”.

Here are the initial images I got (subject to minor refinements of the prompt):

The images are technically quite good, aesthetically quite nice (although it’s a bit of a romance novel cover aesthetic – by the looks of it ChatGPT doesn’t do subtlety, at least not by default), but:

-??????? The chins, the jawlines, the cheekbones, the hair – it’s a hugely exaggerated representation of masculinity. Kitsch aesthetic aside, very few men IRL have such strongly masculine facial features. All these guys looked like they could grace the Abercrombie&Fitch catalogue circa 2003. And it wasn’t just one image, it was all of them.

-??????? All these men were quite obviously Caucasian (white). I intentionally didn’t specify race or nationality in my initial prompt, because I was curious what I’d get, and, sadly, I got exactly what I was expecting. ChatGPT’s “default man” (person?) is white.

?

I was keen on including a diverse array of faces in my PowerPoint, so in the next iteration I asked to generate images of Latino men. Here’s what I got:

OK. As I’m beginning to talk about race and ethnicity, I want to be really mindful. Generally “Latino” refers to people of Latin American ancestry. That covers a broad spectrum of races, including but not limited to Caucasian, Indigenous and Black people. All I got were reasonably light-skinned men with mostly European features.

?

Next, I asked for Asian men:

OK… The three Asian men are undeniably handsome and beautiful, but not one of them has a slim face or one with more delicate features. It is, once again, chin, cheekbones and jawline galore. I know a couple of Asian men who look like that, but… not many.

?

Next up I asked for an image of a Bangladeshi man:

…and then an Indian man:

Next I asked for an image of an Egyptian man – and ChatGPT wasn’t able to produce any image at all. Egypt’s population is approx. 111 – 116 million people. It’s the world’s 14th most populated country. Finding out that ChatGPT was unable to generate an image of an Egyptian man was really baffling.

?

I tried asking for a North African man. Here’s what I got:

Then I asked for a Middle Eastern man and got this:

OK, there seems to a pattern emerging here: no matter the race or ethnicity, I get faces seemingly based on the same visual template in terms of bone structure. It also seems that no matter the race or ethnicity (Black being an exception – more on that below) all the faces are decidedly light-skinned and more often than not look European (sort of vaguely Mediterranean). And what is it with the hair? Very few women I know have locks that thick and wavy, let alone men.

?

Next I asked for an image of a Caribbean man:

Then I asked for an image of a Nigerian man:

And, finally, African American:

I’m an adult; I think I’m realistic when it comes to my expectations of AI in general and gen AI in particular. I am not taking any offence with these images, but I would argue they are slightly problematic nonetheless: they seem to conform to one standard of attractiveness regardless of race or ethnicity. It’s something between “Instagram on steroids”, “shampoo commercial”, “looksmaxxing” and fitness mag cover. I’m extrapolating here, but I strongly doubt that any of these virtual men have more than 10% body fat. They all look like they could bench and squat… A LOT. They all have a somewhat dysmorphic look. There’s not much diversity nor representation in these images. With some adjustments for ethnic features, all these men look kind of the same.

?

Another slightly worrying observation: I didn’t say anything about clothing or lack thereof in my prompts. Since I wanted headshots, I didn’t even think of clothing. However, as ChatGPT was generating the images I couldn’t help noticing that Black and Caribbean men seemed to be shirtless more than all the other guys. So I counted:

-??????? Out of a total of 8 images of Black and Caribbean men, 5 were shirtless (62.5%);

-??????? Out of a total of 21 images of all other men, 7 were shirtless (33.3%).

I don’t know what to make of it. Maybe I’m seeing problems where there aren’t any, but it does feel a bit… odd.

My Gikii slide deck is finished. I'm not sure whether my own dysmorphia has been triggered in any way by the images of those "inhumanly beautiful" men, but I know for a fact that I don't look like any one of them (including "Polish man", which I didn't post above), and neither do most of my male friends of various nationalities and ethnicities. I have great respect for gen AI technology. I just hope that going forward it will become more mindful and nuanced when it comes to broadly defined representation and diversity, because these things matter. They really do.

?