NIST measures demographic gaps across 189 algorithms
On 19 December 2019 the United States National Institute of Standards and Technology published NISTIR 8280 by Patrick Grother, Mei Ngan and Kayee Hanaoka. A total of 18.27 million images of 8.49 million people from four government datasets were run through 189 mostly commercial algorithms from 99 developers. False positive rates across demographic groups often vary by factors of 10 to beyond 100, while false negatives usually vary by factors below 3.
Why it matters
This is the same measurement the outside audits had made, carried out by the body that sets the standards and whose tests vendors sit voluntarily. After it the claim of a demographic gap no longer belonged to the industry's critics: it could be cited from a government report, and it is in that form that it entered the bans and hearings that followed.
False positives: on higher-quality portrait photos the rate is highest in West and East African and East Asian people and lowest in Eastern Europeans, with a factor of 100 between countries. On United States law enforcement images the highest rates are in American Indians, with elevated rates in African American and Asian populations. Rates are higher in women than in men, consistently across algorithms and datasets, although this effect is smaller than the one due to race; they are also elevated in the elderly and in children. False negatives: on domestic mugshots they are higher in Asian and American Indian individuals, and on the lower-quality border crossing images in people born in Africa and the Caribbean, a divergence the report attributes to image quality rather than to the groups themselves. One finding in the report contradicts the simplest explanation: in a number of algorithms developed in China the effect is reversed, with low false positive rates on East Asian faces. The gap therefore tracks the training population rather than being a property of faces. The record does not claim that gender classification was measured: what was tested is one-to-one verification and one-to-many identification, that is, recognising a person.