What Everyone is Getting Wrong about Predictive Policing

by Olivia Zhu

USA Dodge Magnum 01Predictive policing is catching the public’s attention. Interest in the topic hasn’t abated, ever since greater scrutiny, strained budgets, and racial tension have plagued police departments and the communities they are meant to protect. The Marshall Project and ProPublica, among a host of other news organizations, have published in-depth—and extremely popular—descriptions and critiques of the trend.

These pieces merely scratch the surface of the technologies and methods required for predictive policing. The majority of discussions in this space focus on the ethics involved: Are the results increasing instances of racial profiling? Does the practice violate Fourth Amendment rights?

But here’s the thing.

Although predictive policing is in its infancy with regard to adoption and success, there’s far more to it—and there are better questions to be asking.

For example, journalists have wondered about the quality of the data that police departments are shifting into newly purchased software programs. It’s certainly not wrong to state that predictive models will only be as good as the data that serves as their foundation. Nevertheless, assessors have quibbled over whether the data that police departments collect under- or over-represents poor, often minority-dominated communities.

One theory goes that these underserved communities don’t trust the police, and thus are less likely to report crime—making it less likely they’ll be served by any of the benefits of predictive policing. Conversely, perhaps police presuppose that certain neighborhoods are more prone to crime, and decide to patrol them more frequently. That, in turn, increases the likelihood that more incident reports are filed for the region. Predictive models suggest more patrols in these areas, and racial profiling may occur as a result.

The very first set of questions that should be asked, then, is: “How can we determine if under- or overreporting is happening?” “Do reporting trends vary by type of crime?” and “Once we know, can we fold the knowledge into effective predictive policing programs?”

Startups like ShotSpotter and Knightscope might help with the data issue, as they collect and report phenomena such as gunshots, license plates, and other real-time data. Though they may help address the question of how faithful the data is to actual events, other critics have questioned if their kind of data collection may present a privacy concern.

The data that police will consistently rely upon comes from their own troves of paperwork and documentation, in addition to the information collected by private vendors or provided by other jurisdictions. Combining that data when needed is going to be extremely difficult, requiring customization for each department’s idiosyncrasies. Scaling that fusion in a way that is lucrative for the software vendors but still useful for the communities at hand is going to be an interesting problem to tackle, and it seems unlikely to happen unless coupled with massive overhaul of IT infrastructure and thorough redesign of data collection processes.

Thus, the next question set should be: “Are predictive policing programs being implemented in a way that encourages maximum success, given the dollar and time investments of the departments?” “How is success being measured—and are the success rates colored by the fact that the programs themselves changed what data is being collected?”

Moving forward, it’s clear that the most important metrics for stakeholders involve reducing incidents of violent crime. Those statistics are the ones that grab the attention of journalists and citizens alike—they’re sexy, for one thing, beyond being undeniably important.

Yet to focus only on violent crime would be an oversight. The St. Louis County Police Department, for example, had asked a software vendor called HunchLab to only present analyses of violent crimes to mitigate the effects of “racially disproportionate policing.” According to the article, HunchLab and the St. Louis police are seeing some promising results. But it’s important to note that numbers relating to petty crimes (that could still exclude ones excessively tilted to minorities, such as drug arrests) are more robust. The data set is bigger, and therefore could be more reliable.

So, what should police departments be focusing on? First and foremost, I—as a citizen—would expect there to be methods to identify such quotidian, apparently mundane details such as how ticket quotas should be set, or where and when officers could be best deployed to manage traffic flows. The ripple effects of these and other miniscule improvements are what management consultants and manufacturers have been using to make entire industries more productive for years upon years—so why not public sector organizations, too?

Moreover, policing involves so much more than estimating when and where crimes will be committed—the issues that generate so much play in the media and so much concern by civil rights groups. The market has not addressed critical questions of how, say, to improve staffing efficiency, or to manage predicting headcounts given expected population increases or decreases. Even further—could predictive policing help identify problem behavior in certain officers, not just offenders?

The final note is going to seem trite—and yes, I’m warning you in advance. The media needs to stop associating predictive policing with Minority Report, either the short story, film, or short-lived television series. First of all, it’s a lazy analogy to present for a layman reader. More critically, it limits the imagination in thinking of what predictive policing is capable of and should be leveraged for. Good data, math, software, and policies shouldn’t be locked in a black box labeled “Precog”—they need to be understood, criticized, lauded, and used well.