Foreign Study Finds Racial Bias in Facial Recognition Systems

2019-12-23

A recent study released by the US government reveals that many facial recognition systems are more likely to identify people of color compared to white individuals, raising concerns about the widespread use of this technology by law enforcement agencies. The study, conducted by the National Institute of Standards and Technology (NIST), found that when performing a specific type of database search called "one-to-one" matching, facial recognition algorithms frequently misidentify African American and Asian faces at rates 10 to 100 times higher than white faces. The research also discovered that African American women are more prone to being misidentified in "one-to-many" matches, which can have serious consequences in criminal investigations. Despite some companies downplaying early biases in the technology, NIST's study demonstrates that facial matching struggles to overcome demographic disparities. Joey Bramowitz, the founder of Algorithmic Justice League, describes this report as a "comprehensive rebuttal" to the notion that AI bias is no longer a problem. The study comes at a time when the US is growing increasingly dissatisfied with this technology, with critics warning that it could lead to unfair harassment or arrests. In this report, NIST tested 189 algorithms from 99 developers, excluding companies like Amazon that did not submit their algorithms for review. The testing conducted by NIST differs from what companies sell, as it focuses on algorithms that are independent of cloud computing and proprietary training data. The report highlights that SenseTime, a Chinese AI startup valued at over $7.5 billion, had "high rates of false matches" in one of NIST's tests. SenseTime's algorithm produced an error rate of over 10% when examining photos of Somali men. If deployed at airports, this would mean that one in ten Somali individuals could potentially pass through customs using someone else's passport. On the other hand, Yitu, another Chinese AI startup, demonstrated more accurate results with minimal racial bias in its data. In some cases of one-to-many testing, the false positive rate for women of color was nearly ten times higher than that for men of color. In the related one-to-many tests, its algorithm showed minimal differences when comparing photos of black and white men. Microsoft has stated that it is reviewing the report and has not yet provided any comments. Representative Bennie, the Chairman of the House Homeland Security Committee, stated, "Given these alarming findings, the government must reassess its plans for facial recognition technology."

RECENT AI NEWS

RECENT AI TOOLS