A National Institute of Standards and Technology study released December 24, 2019, documents that two facial recognition algorithms misidentify Black women as men nearly 35 percent of the time. The report, a federal government document rather than activist speculation, examined widely deployed systems currently used by federal agencies for border security and criminal monitoring. The study found that error rates spike when the software processes images of non-white individuals.
Asian and African American people were confused with one another up to 100 times more often than white people. These figures reflect real products operating on real data, not hypothetical future systems.
The 35 percent gender misidentification rate for Black females means the algorithm returned the wrong gender more than one in three times. This suggests skewed training data. If the software learned predominantly from images of white men, it will fail on anyone else, according to the data.
False positives present another danger. For Asian and African American individuals, the technology confused two people from those groups at rates up to 100 times higher than for white people.
In law enforcement, a false positive means an innocent person gets flagged as a match, a traveler gets pulled aside at an airport, or a suspect database returns the wrong name. The report serves as a warning. The government itself states that the tools it deploys contain systematic errors that fall hardest on non-white citizens.
Civil liberties are a direct casualty of poorly built software, not an abstract concern. What happens next depends on whether federal agencies act.
The study makes no recommendations. It simply reports the numbers. Those numbers point to a fundamental flaw in how the technology was trained and deployed.
Algorithms learn from data. If the data is not representative, the algorithm is not reliable.
Nearly every system tested showed higher error rates for non-white demographics. This is not a problem one company can fix alone; it is a pattern across the industry. The government is the largest customer for these systems. If federal agencies demand better performance on non-white faces, vendors will respond.
If they continue buying the same flawed products, the errors will persist. The December 24 release date is notable.
Christmas Eve is not a typical day for major government announcements. The findings are buried in a federal study that most Americans will never read. But the numbers are stark enough that civil liberties groups, lawmakers, and journalists will pull them into the open.
The study does not mandate changes. It does not name specific agencies or systems.
It presents the data and leaves the implications hanging. The implications are clear enough: a technology that misidentifies non-white people at dramatically higher rates than white people is not ready for widespread deployment. The government said so itself.
