Fighting Bias in AI

Image generated by MidJourney, Prompt: young man with glasses looking amazmed at the screen of his laptop, the screen is lighting up his face and he is sitting in a dark photo studio.

Talk about a brave new world! As visual communicators, I feel like we are very prone to an emotion or a dot of inspiration that comes from an unlikely place. Much like most of you - the first time I saw generative AI put a teddy bear on a skateboard in Times Square NYC, my brain melted a little bit and I thought, this is going to change everything! Now fast forward 4 years and we’re on the precipice of text to video. Really excited about what we as artists can do with this new tool. As excited as I am about this, there are, and rightfully so, as many people who are legitimately terrified, frustrated, or just flat out angry with what the computers are spitting out. We’ve all seen some pretty amazing visuals, but now people are even winning art contests with AI!

I had my first experience with this a few years ago, I read an article about an artist working with the NFT community. Remember when that was the hot thing? They simply talked about cultural misrepresentation and overt racist bias while minting some NFT on some website that’s lost to the internet - but it made me think. And being the tech nerd that I am, I start looking into how NTFs were created. This led me down a rabbit hole where I first learned about how machine learning worked. How we train computers on specific datasets and how these data set are used to generate images. I thought, “What!? This is wild!” Fast forward a few short year and now we call these lab experiments, generative AI.*

Want to make sure we cover this, but here is a quick overview of how AI learns. All these computer algorithms, programs or “models” are fed data, in our case images and text, and that feeds or “trains” the AI to predict these images for you. It figures all this out based on the correlation of image to word and a very complicated noise reduction model. Now there’s a whole lot more to this but let’s keep it surface level. We feed photos and words to AI, AI uses those as a reference to understand what we mean when we talk to it.

Let’s get into why this is important. Imagine this, all you have ever told your text to imagine an AI machine is that all peppers are long and green. That means every image it will return when you ask for a pepper will be long and green. It will never show you any other type of pepper. It might change some parameters around, but that pepper, but every time you type in “pepper,” a long green pepper will be generated. Because it is all that it “knows” (or was trained on) about peppers. 

“Chat GPT - is trained on a catalog of more than 300 billion words.”

Prompt: a grid of 16 pixel style 8-bit streetwear apes NFTs from 2019

Let’s zoom out a little bit, ChatGPT - I would say our current “a” list celebrity in this space is trained on a catalog of more than 300 billion words, I heard someone call it the “world’s most sophisticated autocomplete system.” It basically just guesses what the next word needs to be based in the data cloud. And because it has so much text, it’s pretty damn good at guessing. 

So, let’s widen that scope, even more at this point billions and billions of images as well as words have now been fed into ChatGPT, MidJourney, DALL-E 3, FireFly (the engine Adobe’s Photoshop AI runs on), and dozens of other AI models. Let’s pause and marvel at this, for a second. Thousands of people worked together to build these systems. It’s incredible, NOW let’s talk about what is being fed into these and what is coming out. At this point most of us have seen the articles where the AI determined that only white and asian men can be scientists and represent women and people of color as working low-paying jobs. If you haven't head over to Google and grab a cup of strong coffee, it’s a wild ride. So, the question is, how did this happen? And what do we do about it?

Prompt: the human condition of bias becoming AI's condition

This data has to come from somewhere, you guessed it, humans. And as we all know how great as a species we all are at knowing and owning our own shortcomings and biases, it’s not hard to imagine that these sneak in when we code the models. It’s something that’s always been part of the human condition and now apparently, we have made it part of the AI’s condition.

And this is not a new thing, I saw an article talking about a UK university admissions program in 1988 that had to be taken offline after they found it to be discriminating against women and people with non-european names. Turns out it was programmed to align with what the admissions department was already doing. Ok, enough examples - let’s talk about how this is bad for art. 

There’s a great NYTimes piece from last year talking about how the generative AI’s are completely misrepresenting specifically black women crying or laughing. It’s worth the read, but I think the real take away here is as artists we are presenting a view of the world that is unique. Sometimes disturbing, sometimes beautiful, sometimes pensive or thoughtful. Remember the main goal of art. Communication. 

One of my favorite pieces by the incomparable Banksy, is a stencil of an unhoused man holding a sign that just says “keep your coins, I want change.” Powerful message, communicated through art. Let’s take this idea to our new medium of AI. 

MidJourney is my current favorite so, I am going to type in a prompt of:

/imagine: unhoused man holding a sign that just says “keep your coins, I want change.”

So, according to this exploration 15 out of 16 unhoused men are over 65 and 10 of 16 are people of color. How many of you believe that? Yeah - go pick a topic and try it for yourself. Is this your vision? Or is it the predictive algorithms?

Moving forward I have a few simple thoughts, you can google “combating bias in AI” and there’s literally hundreds of articles out there, some but, here’s my take away as an artist and my action points as a business owner. None of which require you to be on a board of directors for a think tank, good old fashioned grass roots ideas here.

Be critical - look for patterns and be hard on the art, imagine you’re in a room full of art historians - what would they say? Does this feel like something you would create, if not change your approach, get specific. Remember these AI’s are learning with everything keystroke, we are teaching is as it generating.

Does it pass the jerk test - run a set of prototypes or concepts by your meanest friends, see what they say, do they see patterns? Do they think it lines up with you?

And I think the most important thing to think about it is, YOU are the artist, not the computer. Just as a painter cannot create oil paint from their fingers, that doesn’t make us think less of what they can do with the brush, right? We have used tools to make art since before language was invented, this is just another tool. Lean into it, but remember you, the human, are the one driving, not the machine.

*Machine learning algorithms unearth patterns and generative AI transforms them into something actionable. There’s a solid article here if you want to dive in to this more.

**All images in this article were generated using MidJorney version 6 with various settings.

Prompt: the human condition becoming AI's condition