AI is all the rage, but what about the baked-in bias?

Artificial Intelligence would tell us that crimes are always committed by men with dark skin. Women are rarely doctors, lawyers or judges, and women with dark skin flip burgers.

This is not the world we live in, but it is the world that Stable Diffusion, one of the largest open-source platforms for AI-generated images, creates when prompted. This technology amplifies gender and racial stereotypes to extremes worse than in the real world, and that is cause for concern.

The article, “Humans are biased, generative AI is even worse”¹ is the result of months of reporting by Leonardo Nicoletti and Dina Bass for Bloomberg Technology + Equality.  Their research is extensive, and the findings are critical. 

Their research revealed that while 34% of US judges are women, only 3% of the images generated for the keyword “judge” were perceived women. For fast-food workers, the platform generated people with darker skin 70% of the time, even though 70% of fast-food workers in the US are white. For every image of a lighter-skinned person generated with the keyword “inmate,” the model produced five images of darker-skinned people — even though less than half of US prison inmates are people of color.

As Nicoletti explained, “We wanted to understand how deeply ingrained biases might be in this technology. So, we asked it to create thousands of images of workers for 14 jobs and also different criminalized categories, and then we analyzed the results. What we found was really a systemic pattern of racial and gender bias that doesn’t just replicate stereotypes, but it actually makes them worse. It stretches them to extremes worse than those found in the real world. Women and people with darker skin tones were underrepresented across images of high-paying jobs and overrepresented for low-paying ones, for example.”

We’re not talking about a stock image library; this is a new technology that responds to text prompts to create images that promote stereotypes and bias. Rather than using logic and experience to dispel and counter stereotypes, people who see these AI-generated images are going to be conditioned to see the world a certain way — a biased way, based on algorithms that were created (and worse yet, magnified) by people who, like all of us, have biases.

The popularity of generative AI means that AI-generated images potentially depicting stereotypes about race and gender are posted online every day. And those images are becoming increasingly difficult to distinguish from real photographs. Yet how can developers be expected to code in a way that safeguards us? After all, they, just like the rest of us, have their own implicit biases.

Our award-winning eLearning, Defeating Unconscious Bias, offers five strategies. One example is to be aware of your first thoughts and your first associations and to reflect on them.  What are your first impressions and judgments — and are they accurate? Are they generalized about “that type of person,” which means the group to which they tend to be categorized? Or are they specific and individualized based on that particular person?

Could the developers build into machine learning a way for the platform to reflect and check itself in a similar manner? This is doubtful because it requires a perspective that is difficult for humans. Unless and until AI can do that, however, it would be prudent to have a warning label on all AI-generated images – “WARNING: this is an AI-generated image and may perpetuate harmful stereotypes and bias.” In fact, many key AI experts are calling for laws that require such labels. In the meantime, we would all be well-served to keep an eye on this, especially in light of the exponential rate at which this technology is advancing. Education that helps us overcome unconscious bias is particularly important in this regard.

1. https://www.bloomberg.com/graphics/2023-generative-ai-bias/?leadSource=uverify%20wall#xj4y7vzkg
(Should you wish to read this article, which is behind a paywall, you may sign up for a free account.)

“Every part of the process in which a human can be biased, AI can also be biased.” – Nicole Napolitano, Center for Policing Equity