Study of 189 facial recognition algorithms shows they misidentify people of color more often than white people, confirming facial recognition tech's racial bias
Facial-recognition systems misidentified people of color more often than white people, a landmark federal study released Thursday shows …
Washington PostDrew Harwell
Context & Ripple Effects
The federal benchmark lands on top of a documented pattern rather than opening a new front: an earlier [[a:926589|study found Microsoft, IBM, and Megvii systems erring 21%-35% of the time on dark-skinned women versus under 1% for light-skinned men]], and reporting had already shown the FBI's face databases hold photos of about half of US adults without consent while its matching algorithm skews toward misidentifying black people. What changes with this study is scale and authority — 189 algorithms tested at once, by a federal lab, turning vendor-by-vendor anecdotes into a systemic finding.
It also arrives as deployment is accelerating faster than disclosure: a later [[a:876798|investigation found police in 15 states used facial recognition in over 1,000 cases since 2020 without telling defendants]], meaning the accuracy gap documented here is operating inside criminal investigations, not just lab benchmarks.
First-order effects
Vendors whose algorithms failed the federal benchmark now face procurement-level evidence of bias, not just academic critique — buyers can no longer treat error-rate disparities as contested claims.
Police agencies relying on these systems inherit a documented false-match risk for people of color at the exact moment their caseloads are growing past 1,000 uses across 15 states.
Second-order effects
Government buyers are pushed toward demanding per-demographic error rates in contracts, forcing vendors to retrain and disclose performance slices they previously reported only in aggregate.
The bias finding compounds the consent problem around FBI-scale databases: a system that both collects faces without permission and misidentifies minorities more often gives civil-liberties groups a two-front argument against expansion.
Third-order effects
If demographic error gaps persist across algorithm generations, facial recognition drifts toward the same regulatory treatment as other safety-critical tools — standardized testing before deployment rather than after harm.
The parallel failure in adjacent biometrics, where [[a:951855|speech systems from Amazon, Apple, Google, IBM, and Microsoft misidentify black speakers nearly twice as often as white ones]], points to a structural issue in how training data is assembled across the whole perception-AI industry, not one product category.
The trend: Biometric AI is moving from self-reported accuracy to independently audited, demographically disaggregated benchmarks, with government testing bodies setting the de facto bar for who may sell into law enforcement.
New: Huge federal review of facial-recognition systems finds that white people get more accurate searches than everyone else. With some algorithms, Asian and African American people were *up to 100 times* more likely to be falsely identified than white men https://wapo.st/2tCGkuN
There are two sides to this: - we should not be developing better, more accurate facial recognition to surveil minorities - facial recognition in China works incredibly accurately compared to American tech, so the racism is just really, really obvious here https://twitter.com/...
“The agency did not test systems from Amazon, Apple, Facebook and Google because they did not submit their algorithms for the federal study” https://www.nytimes.com/... via @NYTimes
@jovialjoy @timnitGebru @rajiinio @ClareAngelyn @JayCStanley .@BennieGThompson: The report “shows facial recognition systems are even more unreliable and racially biased than we feared. ... The Administration must reassess its plans for facial recognition technology in light of t…
Facial-recognition technology also had more difficulty identifying women than men. And it falsely identified older adults up to 10 times more than middle-aged adults. https://www.nytimes.com/...
One false match can lead to missed flights, long interrogations, tense police encounters, false arrests, or worse. Even government scientists are now confirming this surveillance technology is flawed and biased. https://twitter.com/...
As with much of the facial recognition stuff, the problem is not “wow these products are so cutting edge,” it's “wow, these products are remarkably bad.” https://twitter.com/...
Oh who ever would have thought that face recognition systems have been trained to be technically racist and sexist, which compounds the sexist, racist ways they are deployed? https://www.washingtonpost.com/ ...
Detroit - a majority black city - purchased facial recognition tech in 2017 for over $1M. After using the tech more than 600x DPD, this fall, couldn't tell us how many arrests it's led to. https://www.wxyz.com/... https://twitter.com/...
This NIST pretty much confirms all the pioneering work done by @jovialjoy @timnitGebru and @rajiinio wrt bias in facial rec. At this point, it would be v diffult to see how vendors could make the case for any LE applications. https://twitter.com/...
“Asian & African American ppl were up to 100 times more likely to be misidentified than white men, depending on the particular algorithm and type of search. Native Americans had the highest false-positive rate of all ethnicities, according to the study.” https://www.washingtonpos…
No surprise here but if we care to preserve any justice in our institutions, this is further warrant for a hard chill on govt and security use of these systems. https://twitter.com/...
1) NIST has a new evaluation of Facial Recognition software out. Not surprising, the software keeps getting better and more accurate but demographics matter. https://www.nist.gov/...
“The majority of face recognition algorithms exhibit demographic differentials. A differential means that an algorithm's ability to match two images of the same person varies from one demographic group to another.” So, maybe not for policing.... https://www.nist.gov/...