By Alex Najibi
We unlock our iPhones with a glance and wonder how Facebook knew to tag us in that photo. But face recognition, the technology behind these features, is more than just a gimmick. It is employed for law enforcement surveillance, airport passenger screening, and employment and housing decisions. Despite widespread adoption, face recognition was recently banned for use by police and local agencies in several cities, including Boston and San Francisco. Why? Of the dominant biometrics in use (fingerprint, iris, palm, voice, and face), face recognition is the least accurate and is rife with privacy concerns.
Police use face recognition to compare suspects’ photos to mugshots and driver’s license images; it is estimated that almost half of American adults – over 117 million people, as of 2016 – have photos within a facial recognition network used by law enforcement. This participation occurs without consent, or even awareness, and is bolstered by a lack of legislative oversight. More disturbingly, however, the current implementation of these technologies involves significant racial bias, particularly against Black Americans. Even if accurate, face recognition empowers a law enforcement system with a long history of racist and anti-activist surveillance and can widen pre-existing inequalities.
Inequity in face recognition algorithms
Face recognition algorithms boast high classification accuracy (over 90%), but these outcomes are not universal. A growing body of research exposes divergent error rates across demographic groups, with the poorest accuracy consistently found in subjects who are female, Black, and 18-30 years old. In the landmark 2018 “Gender Shades” project, an intersectional approach was applied to appraise three gender classification algorithms, including those developed by IBM and Microsoft. Subjects were grouped into four categories: darker-skinned females, darker-skinned males, lighter-skinned females, and lighter-skinned males. All three algorithms performed the worst on darker-skinned females, with error rates up to 34% higher than for lighter-skinned males (Figure 1). Independent assessment by the National Institute of Standards and Technology (NIST) has confirmed these studies, finding that face recognition technologies across 189 algorithms are least accurate on women of color.
These compelling results have prompted immediate responses, shaping an ongoing discourse around equity in face recognition. IBM and Microsoft announced steps to reduce bias by modifying testing cohorts and improving data collection on specific demographics. A Gender Shades re-audit confirmed a decrease in error rates on Black females and investigated more algorithms including Amazon’s Rekognition, which also showed racial bias against darker-skinned women (31% error in gender classification). This result corroborated an earlier assessment of Rekognition’s face-matching capability by the American Civil Liberties Union (ACLU), in which 28 members of Congress, disproportionately people of color, were incorrectly matched with mugshot images. However, Amazon’s responses were defensive, alleging issues with auditors’ methodology rather than addressing racial bias. As Amazon has marketed its technology to law enforcement, these discrepancies are concerning. Companies that provide these services have a responsibility to ensure that they are equitable – both in their technologies and in their applications.
Face recognition in racial discrimination by law enforcement
Another key source of racial discrimination in face recognition lies in its utilization. In 18th century New York, “lantern laws” required enslaved people to carry lanterns after dark to be publicly visible. Advocates fear that even if face recognition algorithms are made equitable, the technologies could be applied with the same spirit, disproportionately harming the Black community in line with existing racist patterns of law enforcement. Additionally, face recognition can potentially target other marginalized populations, such as undocumented immigrants by ICE, or Muslim citizens by the NYPD.
Discriminatory law enforcement practices were highlighted following the murder of George Floyd by the Minneapolis PD. Black Americans are more likely to be arrested and incarcerated for minor crimes than White Americans. Consequently, Black people are overrepresented in mugshot data, which face recognition uses to make predictions. The Black presence in such systems creates a feed-forward loop whereby racist policing strategies lead to disproportionate arrests of Black people, who are then subject to future surveillance. For example, the NYPD maintains a database of 42,000 “gang affiliates” – 99% Black and Latinx – with no requirements to prove suspected gang affiliation. In fact, certain police departments use gang member identification as a productivity measure, incentivizing false reports. For participants, inclusion in these monitoring databases can lead to harsher sentencing and higher bails– or denial of bail altogether.
But how specifically do unjust applications of face recognition and surveillance harm Black Americans? As stated by the Algorithmic Justice League, “face surveillance threatens rights including privacy, freedom of expression, freedom of association and due process.” Surveillance is linked to behavioral changes including self-censorship and avoiding activism for fear of retribution; for example, face recognition was employed to monitor and identify Black Lives Matter protestors. The FBI has a long history of surveilling prominent Black activists and leaders to track and suppress their efforts. Additionally, continual surveillance induces fear and psychological harm, rendering subjects vulnerable to targeted abuses, as well as physical harm, by expanding systems of government oversight used to deny access to healthcare and welfare. In a criminal justice setting, face recognition technologies that are inherently biased in their accuracy can misidentify suspects, incarcerating innocent Black Americans.
In a striking example, the model surveillance program Project Green Light (PGL) was enacted in 2016, installing high-definition cameras throughout the city of Detroit. The data, which streams directly to Detroit PD, can be tested for face recognition against criminal databases, driver’s licenses, and state ID photos; almost every Michigan resident is in this system. But PGL stations are not distributed equally: surveillance correlates with majority-Black areas, avoiding White and Asian enclaves (Figure 2). In interviewing residents, a critical analysis of PGL reported in 2019 that “surveillance and data collection was deeply connected to diversion of public benefits, insecure housing, loss of employment opportunities, and the policing and subsequent criminalization of the community members that come into contact with these surveillance systems.” PGL illustrates how systems of face monitoring can perpetuate racial inequality if their application is not regulated.
Building a more equitable face recognition landscape
Several avenues are being pursued to address these inequities. Some target technical algorithmic performance. First, algorithms can train on diverse and representative datasets, as standard training databases are predominantly White and male. Inclusion within these datasets should require consent by each individual. Second, the data sources (photos) can be made more equitable. Default camera settings are often not optimized to capture darker skin tones, resulting in lower-quality database images of Black Americans. Establishing standards of image quality to run face recognition, and settings for photographing Black subjects, can reduce this effect. Third, to assess performance, regular and ethical auditing, especially considering intersecting identities (i.e. young, darker-skinned, and female, for example), by NIST or other independent sources can hold face recognition companies accountable for remaining methodological biases.
Other approaches target the application setting. Legislation can monitor the use of face recognition technology, as even if face recognition algorithms are made perfectly accurate, their contributions to mass surveillance and selective deployment against racial minorities must be curtailed. Multiple advocacy groups have engaged with lawmakers, educating on racial literacy in face recognition and demanding accountability and transparency from producers. For example, the Safe Face Pledge calls on organizations to address bias in their technologies and evaluate their application. Such efforts have already achieved some progress. The 2019 Algorithmic Accountability Act empowered the Federal Trade Commission to regulate companies, enacting obligations to assess algorithmic training, accuracy, and data privacy. Furthermore, several Congressional hearings have specifically considered anti-Black discrimination in face recognition. The powerful protests following the murder of George Floyd also drove significant change. Congressional Democrats introduced a police reform bill containing stipulations to restrain the use of face recognition technologies. More astonishing was the tech response: IBM discontinued its system, Amazon announced a one-year freeze on police use of Rekognition, and Microsoft halted sales of its face recognition technology to the police until federal regulations are instituted. These advances have supported calls for more progressive legislation, such as the movements to reform or abolish policing. For now, the movement for equitable face recognition is intertwined with the movement for an equitable criminal justice system.
Face recognition remains a powerful technology with significant implications in both criminal justice and everyday life. Less contentious applications of face recognition exist, for example, assistive technology supporting people with visual impairments. While we focus specifically on face recognition in this article, the discussed problems and solutions are part of broader efforts to identify and eliminate inequalities in the fields of artificial intelligence and machine learning. So the next time we unlock our phone, let’s remember that addressing racial bias within face recognition and its applications is necessary to make these algorithms equitable and even more impactful.
Alex Najibi is a 5th-year Ph.D. candidate studying bioengineering at Harvard University’s School of Engineering and Applied Sciences.
For More Information:
- Learn more about facial recognition in the criminal justice system here
- Check out this article discussing how we can hold AI accountable
- Read Georgetown Law’s report on the unregulated, discriminatory use of face recognition by law enforcement in the U.S.
- Take a look at the Gender Shades project’s research on disparities in face recognition accuracy
This article is part of our special edition on science policy and social justice.
38 thoughts on “Racial Discrimination in Face Recognition Technology”
Awesome post, thanks for sharing.
Interesting. This looks super cool. I haven’t read it all yet, but I’ll be back to read the rest of it.
why are you commenting if you didn’t read it?
why are u commenting back over a year after he posted this comment?
Thank u alex najibi for such a great analysis share with us i am never aware about this before yet. This topic should concern for equitable world.
This article is ridiculous and based upon a left-wing fantasy land. It is amazing the excuses that liberals will use for “disenfranchisement”. Please refund my time for reading this nonsense.
Proud boy? Or too difficult an article to understand?
anyone who tells them anything they don’t want to hear is a liberal. if the facts disagree with them, the facts are liberal. so scientific, educational, and media organizations are all part of a huge conspiracy to make them look like ignorant hillbillies. they don’t have to work too hard, imo.
This article has no mention of liberals. Please grow a bigger brain and learn that not everyone is white and this article is pointing out something that effects thousands of people.
Whites ask, “Why is it always about race?” Because…it always is. Even when we don’t want it to be.
Thank u for such a great analysis share with us i am never aware about this before yet. This topic should concern for equitable world.
It’s interesting that the error, or lack of accuracy, is with black women. Is there any scientific theory as to why? I would be interested to read a follow-up article with theories and possible solutions.
3 of the 5 algorithms struggle more to recognize lighter female faces than darker male faces. Each algorithms have really close values of face recognition between lighter faces and darker male faces.
The lowest correct recognition rate is for darker female faces that can be explained with a combination of smoothness of the traits (explaining the difference in correct recognition between male and female lighter faces),lower contrast values in darker faces (the lighter the face is the largest gap of contrast you have: shadows, lips, hairs, mid tones,…) and the fact that darker colors reflects less light which complicates the analysis of the image even more.
Try this exercise: try to get a good picture of an interracial couple (black and white) without using HDR… have fun! You’ll end up with either crushed blacks or burned whites most of the times. But I guess photography is also oppressive.
Actually, it is discriminatory! This article explains it best: https://www.vox.com/2015/9/18/9348821/photography-race-bias.
Thank you for sharing.
Great share! Thank you
thanks for sharing
Awesome! Information. Great work thank you for sharing such useful information. keep it up all the best. I can also refer you to one of the Thermal Sensors & Facial Recognition Solutions in Hyderabad.
Hi, Thanks for sharing this kind of information. this is really very helpful.
I mean sure it may be an issue with picking up lighting but there’s a big difference between having a poor quality photo due to contrast and being arrested for a crime you didn’t commit because facial recognition is inaccurate for the same reason. The point is that law enforcement shouldn’t be using this technology without some major improvements in accuracy.
Interesting post, thanks for sharing.
The error concerning females with darker complexion has to do with an extensive variety of longer hair styles and makeup usage. It has nothing to do with racial inequity. The technology has a difficult time concerning accuracy with these females because most women paint their faces and their hair often obscures many of the facial recognition markers used to detect a person’s ethnicity. The reality is it is also more difficult to see or properly detect facial details of a person of color, particularly at night–this is simply a matter of fact and not mean to be derogatory in any way. I have an olive complexion and people often think I am black when I’m sitting in my car at night. Except I’m not Black, I’m Jewish and Eastern European. If real people cannot detect me accurately, think how much less accurate a computer is going to be. It has nothing to do with racial bias. It has everything to do with technological limitations.
Your point about the importance of light for darker complexions is valid and so is the statement that the issue has more to do with technological limitations. However, if you read the Gender-shades project and the efforts to test these classifiers on very varied data distribution, you might understand the argument better. Your opening statement implies that only black women wear make-up or predominantly wear long hair which is untrue. These algorithms don’t detect facial details by magic. They learn how to from the data they are fed. If white women can be classified correctly, then black women should too. It has everything to do with racial bias if the algorithms are not trained on data that adequately represent various skin tones with or without makeup. You cannot say it isn’t about racial bias if an algorithm can identify white females with makeup but not black females or if it can always correctly identify more white males in any given dataset. Also, what do you mean by “facial markers that detect ethnicity”? The algorithms in question are designed to identify gender or its intersection with skin tone. how do “facial markers of ethnicity” pass as a parameter for gender classification? And please, do point me to the algorithms detecting ethnicity and the standards for the data being used to train them.
Amazing response! Kudos!!!
Why are so many dark black security officer focusing on bi-racial people today when in the past they have been very abusive to bi-racial people and if not abusive they have been in the position to embezzle funds, tamper with DNA records, etc?
Keep in mind that when it came to labeling bi-racial people or abuse the white man usually plays blind, deaf, and dumb.
Also why are dark blacks so focus on our family tree and Genetic Code instead of their own? And why do blacks feel it is important to be the first Black Cinderella or to play Ann Bolyne , Henry the VIII 2nd wife than steadying their own history for wonderful fairy tales or important historic figures? Are you getting the picture. I am a Bi-Racial person according to most and I don’t intend to be a carbon copy of the white man!
thanks for the information
One of the best sites I have ever visited. I hope to get more content like this in the future.
The Detroit graphic and storyline is massively misleading. Project Greenlight was enacted by the city government of Detroit. It did not cover any areas outside the city limits of Detroit. The Demographic map portion of the graphic shows census data from 2010. In 2010 African Americans made up 90% of Detroit’s population. The camera coverage coincides with the African American population because it coincides with Detroit’s city limits not because of some inherent racism. Again a 90% African American city democratically elected a government made up of primarily African Americans to reduce crime in the city limits of Detroit. The city government implemented Project Greenlight in response to the concerns of the African American residents of Detroit who are almost always the victims of criminal activity within the city limits of Detroit.
The predominantly European American areas shown are Dearborn, Melvindale, Highland Park and Hamtramck among others which are all outside the city of limits of Detroit and therefore outside the scope of Project Greenlight. That is why they don’t cameras. If their voters voted for a surveillance system, they would have it to. They did not.
Here is a link to show the city limits of Detroit
I respect everything that you have written in this blog. Please continue to provide wisdom to more people like me.
Was looking for some takes regarding this topic and I found your article quite informative. It has given me a fresh perspective on the topic tackled. Thanks!
The laws of lenses and light are oppressive.
It’s also oppressive that black people can hide better in the dark than white people.
Thanks for the information. There are many issues I never thought. Hope to learn more next time.
any comments about this.,
Very interesting worth the read.
A lot of useful information.