Imagine being handcuffed in front of your neighbors and family for stealing your watch. After spending hours in jail, it turns out that you were identified as the thief by the facial recognition software that the state police used on the footage inside the store. But you haven’t stolen anything. The software pointed the police to the wrong man.
Unfortunately, this is not hypothetical. This happened three years ago to a black father, Robert Williams, who lives in a suburb of Detroit. Sadly, Williams’ story is not a one-off. There was a recent misidentification case in Louisiana where facial recognition technology led to the wrongful arrest of a black man in Georgia for theft of his wallet.
Our research supports concerns that facial recognition technology (FRT) may exacerbate racial inequalities in policing. It turns out that law enforcement using automatic facial recognition is disproportionately arresting black people. This is due to factors such as the lack of black faces in the algorithm’s training data sets, the belief that these programs are credible, and the prejudice tendencies of police officers themselves that exacerbate these problems. It is considered.
No amount of improvement can eliminate the possibility of racial profiling, but we understand the value of automating the time-consuming manual face-matching process. We also recognize that this technology has the potential to improve public safety. However, given the potential harm of this technology, enforceable safeguards are needed to prevent unconstitutional excesses.
FRT is an artificial intelligence powered technology that attempts to ascertain a person’s identity from an image. Algorithms used by law enforcement agencies are typically developed by companies such as Amazon, Clearview AI, and Microsoft, tailoring their systems for different environments. Despite significant improvements in deep learning techniques, federal tests show that most facial recognition algorithms perform poorly at identifying non-white males.
Civil rights activists warn that the technology makes it difficult to distinguish between dark faces and is likely to increase racial profiling and false arrests. In addition, inaccurate identification increases the chances of missing an arrest.
Some government leaders, including New Orleans Mayor LaToya Cantrell, still tout the technology as a crime-solving tool. Some have endorsed the FRT as a much-needed police coverage enhancement that would allow law enforcement agencies to do more with fewer officers as the shortages facing police forces across the country worsen. Such sentiment explains why more than a quarter of his local and state police forces, and almost half of federal law enforcement agencies, regularly access facial recognition systems despite their flaws. may have explained.
This widespread adoption poses a grave threat to our constitutional rights against unlawful search and seizure.
Recognizing the threat to civil liberties, cities such as San Francisco and Boston have banned or restricted government use of this technology. At the federal level, the Biden administration has unveiled his 2022 “AI Bill of Rights Blueprint.” While intended to incorporate civil rights practices in the design and use of AI technology, the principles of the Blueprint are not binding. Additionally, earlier this year Congressional Democrats reintroduced the Facial Recognition and Biometrics Technology Moratorium Act. The bill would suspend the use of FRT by law enforcement until policymakers can develop regulations and standards that balance constitutional concerns and public safety.
The proposed AI Bill of Rights and moratorium is a necessary first step to protect the public from AI and FRT. However, both efforts fall short. The blueprint does not include the use of AI by law enforcement, and the moratorium only limits the use of automatic facial recognition by federal agencies, not local and state governments.
But as the debate rages over the role of facial recognition in public safety, our research and other findings suggest that even error-free software is safe for non-federal use. Unless acted upon, this technology is likely to contribute to unfair law enforcement actions.
First, the concentration of police resources in many black neighborhoods already creates disproportionate contact between black residents and police officers. Against this backdrop, the communities served by FRT-assisted police have been challenged by the demands and time constraints of policing work, plus an almost blind faith in AI that minimizes user discretion, leading to algorithm-assisted We are more vulnerable to execution disparities because the credibility of our decisions is at risk. in decision making.
Police typically use this technology in three ways. One is an on-site inquiry, search of video footage, or real-time scanning of people passing by surveillance cameras to identify those who have been stopped or arrested. Police upload an image, and within seconds the software compares that image to dozens of photos to generate a lineup of potential suspects.
Enforcement decisions are ultimately up to police officers. But people believe AI is infallible and often don’t question its results. Moreover, using an automated tool is much easier than comparing with the naked eye.
AI-powered law enforcement assistance also psychologically alienates police officers from citizens. This removal from the decision-making process allows executives to be disconnected from their actions. Users may also selectively follow computer-generated guidance, preferring advice that aligns with stereotypes, such as advice on black crimes.
There is no hard evidence that FRT improves crime control. Nonetheless, authorities appear willing to tolerate such racist prejudices as the city struggles to curb crime. This makes people vulnerable to having their rights violated.
Gone are the days of blind acceptance of this technology. Software companies and law enforcement agencies must take immediate steps to mitigate the harm of this technology.
For companies, creating reliable facial recognition software starts with a balanced representation among designers. In the US, most software developers are white men. Studies have shown that this software is much better at identifying the race of programmers. Experts believe such findings are largely due to engineers unwittingly imparting a “race bias” to the algorithm.
Racial bias creeps in as designers unconsciously focus on familiar facial features. The resulting algorithms are tested primarily on people of the same race. As such, many U.S.-made algorithms “learn” by seeing more white faces and are therefore less helpful in recognizing people of other races.
Using a diverse training set can reduce bias in FRT performance. The algorithm learns how to compare images by training with a set of photos. A disproportionate representation of white males in the training images produces a skewed algorithm, as blacks are overrepresented in mugshot databases and other image repositories commonly used by law enforcement. As a result, AI is more likely to mark black faces as criminals, making innocent black people more likely to be targeted and arrested.
We believe that companies that manufacture these products need to consider the diversity of their staff and images. However, this does not exempt law enforcement from liability. If this technology is to prevent it from exacerbating racial disparities and leading to rights abuses, law enforcement must seriously consider its methods.
For police leaders, the minimum unified similarity score should be applied to the match. After facial recognition software generates a lineup of potential suspects, it ranks the candidates based on how similar an algorithm determines the images are. Ministries now routinely determine their own similarity score criteria, which some experts argue increases the likelihood of wrongful or false arrests.
Adoption of FRT by law enforcement is inevitable and we recognize its value. But where racial disparities already exist as a result of crackdowns, this technology, without proper regulation and transparency, could exacerbate inequalities like those seen in traffic stops and arrests.
Basically police officers need more training in FRT pitfalls, human prejudices and historical discrimination. In addition to guiding officers to use this technology, police and prosecutors should also disclose that they used automatic facial recognition when seeking warrants.
FRT is not foolproof, but following these guidelines will prevent its use causing unnecessary arrests.
This is an opinion and analysis article and the views expressed by the author are not necessarily those of the author. Scientific American.
