The National Institute of Standards and Technology released a study on December 24, 2019, that measured the performance of facial recognition algorithms already in use by federal agencies. The document is a government publication, not activist commentary, and it evaluates real products operating on real data. One of the most striking findings concerns black women. The software misidentified their gender nearly 35 percent of the time, meaning that in more than one out of three instances the system labeled a black woman as male.
The report characterizes this outcome not as a simple bug but as a design failure rooted in how the technology was trained. Error rates increase sharply when the algorithms process images of non‑white individuals.
Asian and African American subjects were confused with one another up to 100 times more frequently than white subjects were. This disparity appears across virtually every system examined, indicating a widespread pattern rather than an isolated incident. The study also examined false positives, which occur when the technology incorrectly matches two different people.
For Asian and African American individuals, the rate of such mistaken matches was up to 100 times higher than for white individuals. In law‑enforcement or border‑security settings, a false positive can lead to an innocent traveler being pulled aside, an incorrect name appearing in a suspect database, or an innocent person being flagged as a match.
Because federal agencies rely on these tools to strengthen borders and monitor criminal activity, the documented biases have direct operational consequences. The report warns that the technology does not perform equally for all demographic groups; it works best for white people, a result the authors attribute to the composition of the training data. If the data used to develop the algorithms consists mainly of images of white men, the resulting models will be unreliable for anyone else.
The authors note that the bias is not a problem that a single vendor can solve alone. It reflects an industry‑wide trend, and the federal government, as the largest purchaser of facial recognition systems, holds significant leverage.
If agencies demand improved performance on non‑white faces, manufacturers will be incentivized to adjust their products. Continued procurement of the existing flawed systems, however, would allow the errors to persist. While the study supplies the evidence, it does not issue specific recommendations, name particular agencies, or prescribe mandatory changes.
It presents the data and leaves the next steps to policymakers and administrators. The implications, however, are clear: a technology that misidentifies non‑white people at dramatically higher rates than white people is not ready for broad deployment.
The timing of the release—Christmas Eve—is atypical for major government announcements, which may have limited immediate public awareness. Nevertheless, the stark figures are likely to attract attention from civil liberties groups, lawmakers, and journalists who will bring the findings into broader discussion. Public trust in institutions is already fragile, and the report suggests that confidence in facial recognition technology is misplaced when its performance varies so markedly by race and gender.
Whether federal agencies act on the evidence will determine if the current systems continue to be used or if steps are taken to address the underlying training data shortcomings.
