National Institute of Standards and Technology. 2019 FRVT tested 189 algorithms from 99 developers. False positive rates 10–100× higher for Black and Asian faces. Algorithms from Asian countries showed reduced differentials. Training data composition is the primary driver. The name on the training set is the name on the bias.
Appears across 2 pieces (1 defined it · 1 discussed it in essays): when this term was part of the conversation. First surfaces May 2026.
