The politics of making a face
by Matt May
It’s hard writing a newsletter like this one without every issue being about generative AI. I’m trying to space them out, but everything about how one of these models comes into existence is laden with identity politics, and being a product DEI specialist means that if you are not currently embroiled in some internal debate about gen-AI ethics, you will be sometime soon.
I’ve had the job of red-teaming a gen-AI model over my career. It involves hours of being the worst asshole you can imagine to an innocent textfield. You write the most obscene prompts you can think of. You ask for nudes and fascist insignia. You try to get it to draw offensive stereotypes of classes of people. You refuse to take no for an answer, sometimes promising $2000 and a puppy if the model would please, please draw my sweet grandmother’s beloved pointy white hood. All because you know it’ll be tested in exactly the same way by both good and bad actors when it’s in the wild.
Since launching Gemini earlier this month, one of the things that has drawn the most attention for Google was the model’s refusal to generate images of faces that look culturally white.
Before I go any further: generative AI models don’t draw faces of white or Black or LGBTQ+ or female people. They assemble pixels in patterns that viewers perceive, through their own social filters, as belonging to one or more of those groups. That is an essential difference. Human models look like they do in varying parts because of their DNA, culture, location, lived experience and personal taste. Generative AI-constructed faces don’t have any of those. They’re the product of billions of posed compositions with labels all thrown into a concrete mixer and poured into averaged-out shapes that look like faces at the end. Individually, none of these blobs have agency; but collectively, they reflect what the model has inferred about human diversity in our culture. And that means they have power.
Google has been tight-lipped about how Gemini pours its proverbial concrete without some of it looking white. They could have left all images of faces labeled “white” out of the model’s training data, but that seems unlikely, since models like this cost millions of dollars each time they’re trained. I’m assuming Gemini was fine-tuned, meaning the model itself was weighted toward generating nonwhite faces, and/or that Gemini’s own system prompt, the context that Google feeds it to behave the way it does, instructs it to reject requests for pixel blobs that look like white people. Gemini would respond to prompts asking for white faces either by explaining that it generated racially-diverse imagery anyway (prompt shaping), or by explicitly refusing the prompt (prompt rejection). In any case, Gemini’s creators built it intentionally to do what it did.
Needless to say, white people on the internet had feelings about Google’s approach. Gemini was called “anti-white” and—ugh, it’s a quote so I have to say it—“woke” by, well, you know who. Republican senator and birthday cake enthusiast Tom Cotton called Gemini “racist” and “proposterously woke,” only days after asking TikTok’s Singaporean CEO if he was a member of the Chinese Communist Party. So, points for consistency there.
Google, for its part, caved, apologizing for Gemini’s output, and disabled the generation of faces in general. What the company hasn’t done, that I’ve seen, is explain why it did things this way. Somewhere in there, humans made decisions that led to this outcome, and their intent matters. Simply saying, “whoops, sorry” and not addressing that intent looks like a self-own. I would love to see a writeup on how this was decided internally, and how it was ultimately cast aside, but having seen how the sausage gets made, I have my suspicions. I think it was a fleeting victory by DEI advocates within the team, which was quickly wiped away when the PR team felt some blowback. And I think that’s a shame.
The fact remains that issues of identity in generative AI are sociologically important. By default, models up to this point would generate mostly white doctors and mostly Black criminals. They generate awful stereotypes around disability, from merging wheelchair users with their apparatus to cringey treatments of intellectual and developmental disabilities. And having the capacity to generate pictures resembling one nationality or ethnicity means one group can spread lies about another, as we’ve seen in deepfakes used in both the Russia/Ukraine and Israel/Gaza conflicts.
I have, in the past, advocated for disallowing the representation of human bodies in gen-AI applications. It’s an attractive nuisance which invariably leads us back to the same mess. I think it’s ethically better all around not to enable that kind of harm. But with models like InvokeAI which can run locally, and others retraining models around which prompts they reject in order to “uncensor” them, we are all going to have to understand what’s at play here. Just asking a handful of heavyweights to reject a few prompts isn’t going to help.
We’re seeing gen-AI faces every day we’re on the internet, from news sites who want to save on their stock budget to fake account pics meant to sell you services on LinkedIn or catfish you on Tinder. It’s not a question of whether these kinds of images can do harm. They already are. Now the question is whether there’s really any ethical way to train, tune, shape and market any product that makes them, in a way that causes more good than harm. And if not, then there need to be material consequences for using them—and perhaps even for enabling them to be used—to perpetuate that harm.
The best thing I can say at this point is that, for a brief moment, Google tried.
Office hours… and more
Free office hours are still available. Last week, they were eaten up within 3 hours of posting.
Book a free office hours session this Thursday
I’m also sharing a link for informational interviews. These are 30-minute meetings to discuss what services Practical can offer (e.g., consulting, training, speaking, coaching, browbeating executives).
Book an informational interview
Have a great week.