Auditing Computer Vision Beyond Race Labels
A continuous, appearance-based framework for measuring bias in facial recognition systems.
Most audits of facial recognition bias sort faces into a handful of coarse race categories and measure error rates within each bucket. That approach hides a lot — appearance varies continuously, and the categories themselves are doing a lot of unexamined work.
I’m building an alternative: continuous, appearance-based measures that can audit computer vision systems without relying on those labels.
This work is currently in submission as (Bolds et al., 2025).