Artificial Intelligence Predictive Policing: Efficient, or Unfair?
In 2014, 18-year-old Brisha Borden and a friend made an impulse decision to take an unattended scooter and bike, which they immediately returned after the owner of said items showed up. Nevertheless, the girls were charged for burglary and petty theft for the items they stole, worth $80. In the previous year, 41-year-old Vernon Prater, with previous charges of armed robbery and a 5 year sentence in prison, was arrested for shoplifting a Home Depot with goods worth $86.35, similar to Borden.
However, when in jail, Borden, being black, was assigned the label “high risk” for future convictions, and Parker, being white, was assigned “low risk”, despite their criminal histories indicating the opposite. So why were police departments making inaccurate, biased predictions? Were officers making these calls?
Borden and Parker’s risk assessments were done by an Artificial Intelligence algorithm—COMPAS (Correctional Offender Management Profiling for Alternative Sanctions)—meant to predict future crime by convicted individuals, where the results are given to a judge in order to aid them in deciding their sentences. Risk Assessments are part of a larger subset of AI technology being used in Law Enforcement, commonly known as predictive policing.
Predictive policing has been a discussed topic for decades now, but has only been widely implemented in law enforcement relatively recently. The process of predictive policing involves Artificial Intelligence algorithms analyzing large data sets of criminal activity, including arrest and conviction rates as well as demographics of certain areas, in order to determine the at-risk rate for individuals’ future convictions and to determine the concentration of police force in areas based on their rates of crime.
In theory, predictive policing seems like a win-win: avoiding the bias of human error as well as promoting efficiency within Police Departments. But the truth is far from this idealistic take on a technology with the potential to have an extremely large impact on many lives and communities in the US.
Because the AI algorithms base their predictions using historical crime data sets—including data from periods where police departments engaged in unlawful, racially and socioeconomically biased practices—areas that previously have had high rates of crime will automatically be assigned the label “high risk neighborhood”. This leaves such areas unable to reform their image, as being assigned the “crime-ridden” label by the AI algorithm causes the area to be overpoliced, subsequently increasing their rates of arrest and conviction and ultimately perpetuating systemic bias through the unending cycle of policing-arrest-risk. Furthermore, individuals with little previous significant criminal record, like Borden, are assigned the image of “high risk of recidivism” solely due to their racial and socioeconomic background.
Take PredPol, for example. A company based in Santa Cruz, PredPol uses predictive analytics in order to assist law enforcement to predict future criminal activity, essentially dividing cities into sections and assigning them a sort of “crime forecast” in order to determine the concentration of police force in a given area. According to the US Department of Justice Figures, black people are five times as likely to be arrested than white people. This means that the data that the PredPol algorithms draw upon are biased regardless of their supposed impartiality, defeating a major purpose of using such technology: to avoid human partisanship.
The accuracy of predictive policing is not limited to only systemic bias. For instance, if an area had an unusually large rate of crime for a single day, such as a mass murder, then the area would be assigned as “high risk”, causing law enforcement to increase the magnitude of police force in said area. With more officers monitoring citizens, more are bound to be arrested, thus cementing the area into the “high risk” image, showing how one instance can have a permanent impact on a location if predictive policing continues to be heavily utilized by law enforcement.
The AI cannot determine on its own what is biased/inaccurate; it can only interpret and spit out what data was fed to it. With a lack of historically unbiased data, the possibility of an unbiased predictive policing scheme seems unlikely in the future.
So is the technology salvageable? Debatable. To decide upon the future of people’s lives and communities as a whole based off of the judgement of biased algorithms does not seem fair, yet many police departments still use the software. Perhaps in the coming decades, the data that the AI use may be less skewed as it is today, but to think the technology will be able to account for all of society’s nuances is not a feasible expectation, but only time will tell.
“Predictive policing algorithms are racist. They need to be dismantled.”, MIT Technology Review, Will Douglas Heaven, 17 Jul. 2020
“Why Hundreds of Mathematicians are Boycotting Predictive Policing”, Popular Mechanics, Courtney Linder, 20 Jul. 2020
“Machine Bias”, ProPublica, Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, 23 May 2016
“Predictive policing is a scam that perpetuates systemic bias”, The Next Web, Tristan Greene, 21 Feb. 2019