Can AI Help Us See Our Own Bias?

"We like to think we can tell what's real. We trust our eyes to distinguish a human face from a digital one, a photograph from a fabrication. But what happens when we can't? In my recent study, I explored how AI-generated faces can be used to probe one of the most pervasive and least visible forms of prejudice: weight bias. And in doing so, the study exposed something more fundamental."

"When we look at a face, we don't only see it: we judge it. Often, without even knowing we're doing so. It's easy to ask people how they feel about others, much harder to measure what they won't (or can't) admit. That's why researchers often turn to tools like the Implicit Association Test (IAT), which reveals deeper bias by tracking how quickly people associate certain traits with certain groups."

"These flaws introduce noise, such as biases not about weight, but about photographic quality, ethnicity, age, or even expression. And when studying implicit bias, that kind of contamination is fatal. Our approach is simple: using generative AI, we created 48 standardized portraits, each one composed and standardized to depict either an individual of average weight or one with obesity, matched in lighting, pose, background, and expression."

AI-generated faces enable controlled probing of weight bias by creating standardized portraits matched on lighting, pose, background, and expression. Forty-eight portraits representing average weight or obesity were generated and rated by over 200 participants for realism, competence, and perceived weight. Heavier faces received lower competence ratings and were judged less realistic. Traditional weight-bias tests relied on non-diverse, poor-quality photographs that introduced confounds such as photographic quality, ethnicity, age, and expression. Standardized AI images remove many visual confounds, improving the precision of implicit bias measurement and offering a tool for future bias-reduction research.

Read at Psychology Today