Gizmodo published an article yesterday titled, AI Image Generators Routinely Display Gender and Cultural Bias, in which Kyle Barr argues that the results produced by AI image generators do not reflect the real distribution of images on the internet.
The author makes a flawed and irrelevant comparison:
“The U.S. Bureau of Labor Statistics shows that women are massively underrepresented in the engineering field, but averages from 2018 show that women make up around a fifth of people in engineering professions. But if you use Stable Diffusion to display an ‘engineer’ all of them are men. If Stable Diffusion matched reality, then out of nine images based on a prompt ‘engineer,’ 1.8 of those images should display women.
First of all, let’s note the Americentric bias that Mr. Barr himself demonstrates. Believe it or not, there are many engineers outside the US, and many pictures of them on the internet. Perhaps the distribution outside the US is different?
Stable Diffusion and other AI image generators do “match reality.” It’s impossible for them not to. The math is indifferent. If there is perceived bias in the output, there must be bias in the input, the training set. But don’t forget that there is a critical element in the training set in addition to the image itself. Without descriptions of the images, the technology could not exist. There would be no way for the AI to associate any words with any images at all.
The output of the AI should not be expected to match a simple statistical distribution of images on the internet, because that is not what it was designed to do. Understanding the impossibility of bias in the mathematics should prompt you to look deeper. If the output distribution does not match an unbiased measurement of the distribution of image content at large, then the bias must be in the image descriptions, which express sentiment. Big surprise—the bias there is attributable to humans as usual.
You cannot get an unbiased training set. You would have to cherry-pick your inputs, and then you’d be incorporating the sentiment and bias of the cherry-picker.
Knowing that there is an inherent human bias in the AI, “be the change you want to see in the world.” Compensate for it in your prompts. There’s nothing stopping you from asking the AI for a Black female surgeon or a Native American male nanny. As always, the responsibility for combatting bias lies with individuals. We won’t see a general change until more individuals decide to produce less-biased media.