causal

Facial Recognition Fails On Black Women Due To Training Data

Off-the-shelf facial recognition systems fail to detect dark-skinned faces (as Joy Buolamwini found, only recognized when she wore a white mask) not because engineers are racist but because benchmark training data was roughly 88% 'pale and male'—drawn from celebrity and political profiles—so the systems work well for white men and badly for Black women; the algorithm isn't biased so much as we are, garbage in, garbage out.

causalpending

Speaker

Catherine D'Ignazio

Evidence Quote

“a system that's trained with that kind of training data, is gonna fail really badly. It works great for White men... and then it works really terribly for Black women because there are so few images of Black women in the training data”

Source

156 | Catherine D'Ignazio on Data, Objectivity, and Bias— Sean Carroll's Mindscape
Created: 6/13/2026, 12:14:23 AM

Your notes

Loading notes...