Press "Enter" to skip to content

Addressing Racial And Gender Bias In Facial Recognition Surveillance

The dystopian surveillance state of science fiction media is within reach—and some privacy activists argue that it’s already here. Facial recognition advancements have spiked fear and uncertainty over misuse and civil liberties infringements, but with the alarm comes a wave of activists bringing solutions.

What is facial recognition?

Facial recognition is a form of artificial intelligence. Artificial intelligence broadly refers to the development of computers to perform tasks that would normally require human intelligence. If you have an email, you are indebted to AI for directing spam to a separate folder instead of flooding your inbox—the computer learned to recognize the pattern of spam and filter it accordingly. If you have a YouTube account or music streaming service, your personalized recommendations are the product of an AI algorithm.

For facial recognition, algorithms are written to measure the geometry of someone’s face, compare those unique measurements to a database of faces, and return potential matches with varying degrees of certainty.

Facial recognition can offer convenience, such as the millions of people who use their face to unlock their iPhones. It can also be used as a surveillance tool, like the Chinese government using the technology to track Uighurs, a largely Muslim minority the government has been sending to detention camps.

As AI embeds itself into daily technological life, privacy activists and technology enthusiasts agree that the powerful tool is here to stay. But, the implementation of the technology has complex problems, as well as disagreements about which solutions are actually solutions.

Resolving the bias in the technology

The current facial recognition technology available has racial and gender bias. A 2018 study by MIT found that while determining gender using three different facial recognition programs, the error rate for light-skinned men was 0.8%, while darker-skinned women were misgendered 20% to 34% of the time.

In 2018, Amazon’s facial recognition tool, Amazon Rekognition, misidentified 28 members of Congress as people who had been arrested for crimes. While comparing the lawmaker’s faces against mugshot databases, Rekognition misidentified lawmakers of color at a rate of 39% even though they only made up 20% of the people.

Brian Brackeen, founder of facial recognition technology company Kairos, says that it’s possible to train the bias out of the technology.


Facial recognition algorithms improve by being supplied large datasets of faces. Existing facial recognition products work well on “pale males” because the algorithms were supplied datasets of majority White men, reflective of the tech industry itself. An MIT and Stanford University study found that a widely used training set was more than 77% men and 83% White people. Brackeen says this bias toward the people who make the technology is common; similarly, algorithms in Asia tend to perform well on Asian males and not as well on White males.

To solve the bias, the algorithms need to be supplied a more diverse set of faces to learn from, but high-quality datasets of people of color aren’t always available. For example, cellphone photos of dark-skinned people are unreliable because the light sensors are inconsistent and rarely accurately reflect the person’s skin tone, Brackeen says. Additionally, many companies struggle to find enough high-resolution photos that can be consensually added to training datasets—Microsoft had to remove its dataset of more than 10 million images in 2019 because it was full of faces scraped from the web and videos under the Creative Commons license.

That’s why Brackeen supports the use of Generative Adversarial Networks to train facial recognition algorithms. GANs are able to create computer-generated faces made from a composite of photos of real people. The faces are indistinguishable from real photos and can be generated to meet specifications such as age, race, gender, and photo quality.

“We can create a million Black men, a million Asian men, a million whatever, and just load the AI with those generated people,” Brackeen says. Using that method, he believes “there will be no bias in facial recognition AI three years from now.”

But even if facial recognition AI becomes a completely accurate tool, Brackeen believes the technology should never be used by law enforcement. He sees facial recognition as a way to boost convenience. For example, being used as a way to checkout at the grocery store faster—think Amazon Go stores—or being implemented at Disneyland to speed up the wait times on buying passes or getting on rides. In those instances, the consumer would opt-in to using facial recognition and the consequences of the technology failing would be minor annoyances, such as a longer checkout process or slower lines to get on rides.

Government use of the technology

In the hands of law enforcement, however, the tool could prove dangerous. If someone who is pulled over for speeding is misidentified as being on a dangerous persons database via an officer’s body cam feed, that situation could turn fatal.

“These impacts are too great,” Brackeen says.

It’s unclear how many law enforcement departments have access to facial recognition technology and how those departments that do have access use the tool. The Georgetown Law Center on Privacy and Technology was the first group to present a comprehensive report on the status of facial recognition within U.S. police departments in 2016. The team submitted public records requests to more than 100 law enforcement departments, finding that there was no regulation on the facial recognition technology that affects half of Americans.

“We ourselves were surprised by some of the findings we had, mostly in terms of the sheer scope of use of facial recognition at the state and local level across the country,” says Clare Garvie, a senior associate with the Georgetown Law Center on Privacy and Technology. “Also, [we were surprised] at the complete absence of any legal imposed regulations or even policies implemented by law enforcement to constrain its use.”

Garvie’s work revealed that at least 52 police departments of the 100 queried have access to, use, or have previously used facial recognition technology. Several large police departments, including those in Los Angeles, Chicago, and Dallas, either had access to or were exploring the use of real-time facial recognition that can continuously scan pedestrian’s faces and compare them against databases such as mugshots or driver’s licenses.

Additionally, Georgetown’s research found that because of the lack of policy and limitations on how the technology is used, police departments were submitting forensic sketches, digitally altered images, and photos of celebrities who looked like the suspects to the facial recognition algorithms and earnestly using the output.

For example, in 2017, a pixelated image from a surveillance camera of a suspect stealing beer from a CVS returned no matches when run through the NYPD’s facial recognition technology. Someone mentioned that the suspect resembled actor Woody Harrelson, so a picture of Harrelson was ran through the program and the detectives picked someone who they believed looked like the suspect from the resulting matches. That celebrity “match” was sent back to the investigating officers, and an arrest was made.

The stakes are too high in criminal investigations for the police to be using probe photos and doppelgängers, Garvie says.

Additionally, because most facial recognition programs used by law enforcement search mugshot databases, people of color, particularly young Black men, are overrepresented in the possible matches. Even if the technology’s racial bias problem is solved, the use of facial recognition within the existing criminal justice system could just replicate the over-policing of Black and Brown communities.

“It’s not appropriate to ask individuals to affirmatively fight for our right to privacy against government intrusion,” Garvie says. “That said, I think the single most important role individuals can play is demanding transparency and accountability from their local and state officials, and pushing their legislators for the legislation and regulation that they deem appropriate in this field.”

Garvie advocates for a moratorium, or a temporary halt, on government’s use of facial recognition technology while researchers and lawmakers have time to catch up and consider the impacts of this fast-moving technology. San Francisco was the first city in the U.S. to enact a three-year moratorium on facial recognition being used by city government in public spaces.

Immediately after the ordinance was put in place, city employees scrambled to remove the face unlock feature from their government-issued iPhones. Having to remove face unlock was a minor oversight, but Garvie highlights it as an example of why it’s so difficult to develop legislation that anticipates for the future without being too broad and cutting off AI uses that are more helpful than hurtful.

“It’s very challenging to think about legislation that regulates exactly what you want it to, that doesn’t leave loopholes but isn’t also overinclusive,” Garvie says. “It’s a challenge, and it’s never going to be perfect.”

How the moratorium will affect the San Francisco Police Department’s criminal investigations, though, is uncertain. While the SFPD did not respond to YES! Magazine’s request for comment, the Georgetown Law report found the department had been using facial recognition since 2010 and could search a half-million to a million mug shots.

The San Francisco Sheriff’s Department stated that the ban would not affect its work because the department conducts business in nonpublic spaces—such as the county jail—which are not covered by the ordinance.

But how the technology is being used in those nonpublic spaces is unknown.

“I don’t know to what extent we use it,” says Nancy Crowley, director of communications for the sheriff’s department.

Activists such as Evan Greer, director of operations at Fight for the Future, argue that the moratorium in San Francisco is not enough.

“We don’t think there are meaningful limitations you can put on this technology,” says Greer, pointing out the loopholes in the San Francisco ordinance that allow facial recognition to be used in nonpublic spaces.

Greer believes that while developing the language around legislative regulation may be tricky, an outright ban on facial recognition surveillance by government, law enforcement, and private corporations sidesteps the concern for future-proofing and overinclusion.

Fight for the Future, a nonprofit grassroots advocacy group, has built a campaign to ban facial recognition over the past year, making headlines when they got 40 major music festivals including Coachella and SXSW to commit to not using facial recognition at the festivals. The group has since expanded to organizing on college campuses and supporting movements across the nations by acting as an online hub for the disparate movements to ban facial recognition across the nation.

Greer and Fight for the Future’s focus is on slowing down the rapid adoption of facial recognition and giving lawmakers time to do their jobs—put a policy in place that regulates its use. In the meantime, Greer’s opposition to facial recognition surveillance is not quelled by the removal of bias and increasing accuracy of the technology.

“If we have systems in place that make it possible to enforce laws 100% of the time, then there is no space for us to test whether those laws are just,” said Greer, highlighting civil rights, LGBTQ rights, and the legalization of marijuana as examples. “For me, at a philosophical level, privacy isn’t about what you have to hide, it’s about our ability to evolve as a human society.”

Increasing the accessibility of advocacy

Because facial recognition is developed by small groups of people within the tech field, the technology operates within a black box. Yet, its impact is largely on the public, who can’t see inside the black box and who don’t have the information to fully understand its impact.

The end of the 2016 Georgetown Law report offers recommendations for legislatures, law enforcement, facial recognition companies, and community leaders. While suggestions for law enforcement and lawmakers are pretty standard—such as officers must have probable cause to run a photo through facial recognition—the section for community leaders is unique because it provides individuals with the questions they need to ask to get information from their police department. Questions such as: Who is enrolled in the police face recognition database? What legal requirements must be met before officers run a face recognition search? How does the agency’s face recognition use policy protect free speech?

These questions reveal the need to involve people at every level. The Georgetown report reflects only 100 of the 18,000 police departments across the country. Garvie believes getting more people prepared to advocate for themselves is critical to uncovering what we still don’t know about law enforcement’s use of facial recognition.

Preparing people to advocate for themselves is also why Brooklyn-based artist, researcher, and technologist Mimi Onuoha teamed up with Mother Cyborg, aka Detroit-based artist, DJ, and educator Diana Nucera, to make an educational zine about AI.

The two met at a conference about the future of AI, engaging in conversation surrounding the ethics and potential pitfalls of AI. They bonded over how they were going to share this information with their communities. In Detroit, where Nucera lives, 40% of people don’t have access to the internet, so how is she to engage her community in complex AI topics when they are still working on getting them online?

“It was clear to me that we were in this room of pretty privileged folks in the field, talking about these things that really my neighbors should be talking about,” Nucera says.

The two collaborated to distill the most essential concepts of AI down to a simple format that could be used for self-teaching, or as a tool for a teacher to use in the classroom. The zine not only gives communities the tools to understand and talk about AI, but also to foster conversation around the growing tech that isn’t dominated by a fatalist point of view.

“There must be a way to talk about this that doesn’t end in this strange ‘robots are going to kill us’ way,” Onuoha says. “We know that there are alternate conversations we can have.”

The 84-page zine engages in that alternate conversation by including exercises and prompts that demystify algorithm building by having the reader identify algorithms in their own lives. The format is digestible, building upon a basis of knowledge to get to the more difficult ideas, like whether algorithms can be used to quantify and judge people in an ethical way.

The booklet sold out of physical copies twice, and Onuoha and Nucera continue to run workshops for non-tech and tech people alike, leading critical discussion on what regulation and data collection should look like and who is being served or harmed by the use of AI and facial recognition.

Ultimately, the zine provides enough information to act as a launchpad to engage in those critical conversations. Both Onuoha and Nucera want to make sure that their neighbors, mail carriers, and anyone else they may see in their community have the language and information they need to participate in the discussion about AI surveillance being implemented in their community.

“They are the only ones who can say how it’s impacting them and whether or not they want it,” Nucera says. “The rest of it is just assumption. So how do you get people to join those conversations? They’ve got to learn the talk.”

Isabella Garcia is a solutions journalism intern for YES! .
Connect: Twitter