Department of
Psychological and Brain Sciences

In a series of preliminary studies, Hugenberg and his team have shown that, as Hugenberg explains, “racially prejudiced participants tend to use these race-signifying facial structures to judge trustworthiness, conflating personality traits and characteristics of race. They even use these prototypical cues to make inferences about some white people who have less prototypically white characteristics. Trustworthiness and whiteness are fused together in their mind. Less prejudiced participants, by contrast, don’t conflate race and trustworthiness, at least not to the same extent.”

To identify this “race-trait overlap,” the researchers have developed a sophisticated method to empirically measure the overlap between face traits and racial stereotypes. With the use of three-dimensional computer modeling software, participants from a variety of racial groups are shown a series of faces, which vary randomly according to lightness of skin, nose and brow structures, lip thickness and other features. They then ask the participants to make judgements of these faces, for example, how trustworthy they look.

After participants make judgments on hundreds of examples, the researchers are then able to compute what features the participants are using to make their judgments and how strongly they are using one feature or another to make their judgment. They can then pose questions about racial prototypicality of the faces to determine whether these statistical models are aligned and the extent to which participants are using the same underlying facial features to judge personality traits.

Exposing the racism that informs facial perceptions, says Hugenberg, can ultimately make way for some exciting interventions that will reverse the confounding of personality traits with racial prototypes. “If people are using racialized cues and faces to make judgements about trustworthiness, we should be able to mathematically take out the effects of race from trustworthiness and trustworthiness from race. We might be able to retrain them on stimuli that have taken out what they think is there and show them what whiteness or blackness look like to them.”

He sees other exciting possibilities around questions of algorithmic bias. As Hugenberg explains, machine learning algorithms trained to detect facial trustworthiness have (unintentionally) ended up with racial biases. Major insurance companies, for example, have been using facial analytics to predict the personality traits of their customers, thus creating the same possibilities for race-trait overlaps. “Our research,” says Hugenberg, “can not only model the magnitude of these overlaps, but potentially de-bias the statistical models used by such insurers by de-confounding race from other predictors. By computationally decoupling race from facial trustworthiness, this work could help de-bias machine learning algorithms as well.”

Dismantling racism often seems like a daunting and colossal task; yet such an approach suggests that if we can detect the critical points at which it enters our everyday social perceptions, we can expose it – even dismantle it – for what it is.

And just for the record: exactly how trustworthy is our perception of trustworthiness based on facial perception? According to Hugenberg on the existing science, “it’s not trustworthy at all. There’s no accuracy to judging people’s trustworthiness at zero acquaintance. Trustworthiness is an invisible cue. We really need to get behavioral information about a person before we can make a meaningful judgement.”