Stopping Crime Before It Starts
Predictive policing helps police protect citizens. It could also be used to oppress them.
In the 2002 sci-fi thriller Minority Report (based on a dark Philip K. Dick tale), Pre-Crime Unit Captain John Anderton is on the run from police because the mutant pre-cog psychics used by his unit predict that he will murder a man in the next 36 hours. More recently, the hit CBS television series Person of Interest posits a secret all-seeing computer surveillance system developed by a reclusive billionaire genius for the U.S. government that can predict that a specific person will be involved in a violent crime. For now, these are fiction. Researchers are, however, claiming to have developed computer programs that can predict not who will commit a crime, but at what locations they are likely to occur. Welcome to the brave new world of predictive policing.
Predictive policing goes beyond the celebrated CompStat system that was widely adopted by many cities as a national crime wave crested in the 1990s. In CompStat recent crime data are plotted on a city map as a way to identify crime "hotspots" to which more police are deployed. CompStat is credited with dramatically reducing crime in cities where it was implemented.
Now comes predictive policing, which proponents claim is even more effective in reducing crime. For example, the Los Angeles Police Department (LAPD) recently reported in a randomized controlled trial that one such program has reduced property crimes by 13 percent in precincts where it has been implemented compared to a slight overall increase in those crimes in the rest of the city. "We have prevented hundreds and hundreds of people coming home and seeing their homes robbed," said police LAPD Capt. Sean Malinowski to the AP. Malinowski is the Commanding Officer of Real-time Analysis and Critical Response of the LAPD and the principal investigator on the National Institute of Justice funded "Los Angeles Predictive Policing Planning Project." The LAPD is now rolling out the program to more of the city.
The particular program used by the LAPD is called PredPol, which has been devised by a team of researchers led by University of California, Los Angeles anthropologist Jeffrey Brantingham. PredPol is relatively new; other cities have been using predictive crime analytics programs by IBM for several years. In addition, using PredPol the City of Santa Cruz experienced a reduction in burglaries by 27 percent in July of 2011 compared with July 2010. PredPol's algorithm was able to predict crime time and location (hotspots) in Los Angeles with twice the accuracy of trained crime analysts.
In Los Angeles, predictive policing is currently applied to forecasting the likelihood of burglaries, auto theft, and theft from autos. Crunching weighted crime data from the past three years, the PredPol algorithm produces a daily list of hotspot boxes measuring 500 feet by 500 feet and along with times when the crimes are predicted to be most likely to take place. Between responding to specific calls for assistance from the public, officers are directed to go into the boxes identified by the program. The idea is not to make arrests but to disrupt law-breaking before it occurs.
How does predictive policing work to reduce property crimes? Crime does not randomly disperse through cities. For example, research has shown that half the crime in Seattle occurs on 4.5 percent of that city's streets; just over 3 percent of street addresses and intersections generated half the crimes in Minneapolis; and 8 percent of street blocks accounted for 66 percent of robberies in Boston.
Researchers have developed two theories for why some areas are subject to higher rates of crime; near repeat theory and risk terrain modeling. Near repeat theory hypothesizes that once a particular location has been hit by a crime it is more likely nearby locations will be hit too. For example, studies have shown that burglaries are "contagious." One study found that "houses within 200 meters of a burgled home were at an elevated risk of burglary for a period of at least two weeks." Why? Possibly because a successful burglary advertises similar vulnerabilities in other properties in a neighborhood.
Risk terrain modeling maps various risk factors to identify areas where crimes are more likely to occur. For example, Rutgers University computational criminologist Joel Caplan mapped for Irvington, New Jersey four crime risk factors correlated with shooting incidents. The risk factors were the locations of gang member residences, public bus stops, schools, and facilities like bars, clubs, fast food restaurants, and liquor stores. He found that "the likelihood of a shooting happening at particular 100-foot-by-100-foot places in Irvington during 2007 increases by 143 percent as each additional risk factor affects that place."
In June, Brantingham and his colleagues published a study that applied Lotka-Volterra equations used by biologists for decades to determine the hunting ranges of animals in the wild to map the territories of street gangs [PDF]. Their model predicted that 59 percent of gang crimes would occur within two blocks of a border between two gangs and 87.5 percent would occur within about three blocks. When the researchers mapped more than 500 crimes attributed to 13 gangs in a specific area of Los Angeles, they found in fact that 58 percent and 83 percent occurred within two blocks and three blocks of a border respectively.
"You would think that we're more complicated than other animals, so a model this simplistic shouldn't work, but I was surprised that it fit as well as it did," said study co-author Martin Short, an assistant adjunct professor of mathematics at UCLA in Wired UK. This research may eventually be used to identify zones to be more intensively patrolled by police with the goal of disrupting assaults and murders perpetrated by gangs.
The accuracy of predictive policing programs depends on the accuracy of the information they are fed. Lots of crimes are not reported, skewing the computer forecasts. According to near repeat theory, the probability that a crime will be committed decays with time, so moves toward real-time data input would boost the accuracy of forecasts. Finally, it is early days yet for predictive policing so the algorithms must be validated by outside experts. Some police departments have become notorious for cooking their crime statistics books, so an independent oversight board is a good idea to keep them honest. One possible downside of transparency is that savvy criminals or terrorists could use predictive policing programs to figure out likely police deployments as a way identify unprotected targets.
How might predictive policing interfere with the Constitution's Fourth Amendment guarantee that Americans are to be free unreasonable searches and seizures? Andrew Guthrie Ferguson, a law professor at the University of the District of Columbia notes in an article, "Predictive Policing: The Future of Reasonable Suspicion," forthcoming in the Emory Law Journal, that police must have either "probable cause" to search or "reasonable suspicion" to seize an individual. Such determinations are actually predictions by law enforcement officials about the likelihood they will find evidence of a crime when they search a premise or detain a suspect. Can computer programs improve these predictions and thus help police identify would-be perpetrators while excluding the innocent?
To find analogies to how predictive policing might affect Fourth Amendment protections, Ferguson reviews various Fourth Amendment court cases involving anonymous tips, informant tips, profiling, and high crime area designations. Tips refer to the activities of particular individuals. Predictive policing forecasts do not. Consequently, Ferguson argues, "Because predictive policing does not provide personal knowledge about an ongoing crime, or particularized identification of the suspect involved, it cannot support the weight of reasonable suspicion."
On the other hand, if a specific area has been identified by the computer program as being at higher risk for an outbreak of, say, burglaries, then courts would likely accept reasonable suspicion arguments by police who had stopped a suspect in that area fitting a burglar "profile," e.g., carrying duffel bags, tools, ropes, gloves in warm weather, etc. Ferguson concludes that "predictive policing forecasts, alone, will not constitute sufficient information to justify reasonable suspicion or probable cause," but instead will be seen by courts as a "plus factor" in making such determinations.
Ferguson also expresses the hope that the advent of predictive policing might "cause courts to rethink the current overly flexible approach to reasonable suspicion." One possible liberty-enhancing benefit from predictive policing might be that by focusing law enforcement attention on specific city blocks that innocent citizens living in higher crime neighborhoods (often inhabited by members of ethnic minorities) may experience less intrusive police contact. Is it too much to hope that better crime forecasts will not only lead to fewer crimes, but also to less police interference with our liberties? Maybe not. But we should always keep in mind that any new technology that helps the police to better protect citizens can also be used to better oppress them.
Science Correspondent Ronald Bailey is the author of Liberation Biology (Prometheus).