AI face bias beats humans, study finds

The study tested whether AI would escape a well-documented human habit: judging people by their faces. Humans tend to read competence or trustworthiness from facial structure even though it does not reflect character, a bias rooted from childhood. Researchers led by Stephen Lehr of Cangrade, an AI recruiting-tool developer, showed AI models pairs of faces tuned to look more or less competent and more or less trustworthy.

The results ran opposite to the hope that AI could remove such bias. GPT-4o leaned toward the faces humans read as competent far more often than humans did themselves. When the team tested monkeys' faces that humans had rated as kind or mean, GPT-4o still picked the kindly rated ones 66% of the time, hinting it had formed a general concept of facial trustworthiness rather than memorizing human faces. Asked which face was more likely to be a serial killer, human trafficker or financial fraudster, GPT-4o chose the less trustworthy-looking face 68.7% of the time, and it picked the more competent-looking face 75.19% of the time in realistic choices such as selecting a university president or a startup to invest in.

The finding that newer models showed more bias, not less, suggests training data alone does not explain it, the team argues — the models appear to amplify it. The stakes hinge on whether this behavior holds outside forced two-face comparisons and on flat images, and for whom: recruiters and legal teams weighing AI tools are the ones most exposed, the authors suggest.