The Paradoxes of Predictive Policing

• In Event: Politics, Ethics and AI Technology: The State of the Discipline

Abstract

Predictive policing has come under fire for treating arrest data as a suitable proxy for crime data, despite the fact that establishing the incidence of crime through police actions or selective citizen reporting is likely to be racially biased. Lum and Isaac (2016) have shown that doing so leads to pernicious feedback loops that further exaggerate the over-policing of minority communities. Yet this case raises even more fundamental questions about the objective that predictions of crime are meant to serve. Pinning down the goal of predictive policing can be surprisingly difficult. Treating good policing as a matter of predicting where crime might occur in the future suggests that the goal of policing should be for police to observe crime. In other words, predictive policing makes police better at their jobs by ensuring that they are more likely to observe crime whenever they are on patrol. But this could easily generate perverse results: sending police to address predicted crimes might increase the observed crime rate because police simply become more likely to detect it (even if the true crime rate (which includes unobserved crime) has remained unchanged).

A more obvious goal for policing is to reduce the incidence of crime, but it is not obvious how predictions about crime—using arrest data—can help achieve that goal. The only way such predictions would help reduce even the observed crime rate is if the presence of police discourages law-breaking, rather than merely increasing its observation. To decrease the true crime rate, it would also need to prevent its relocation outside the police’s view. In other words, reducing the crime rate would require some way of predicting when the presence of police prevents a crime from occurring. But using arrest data as crime data makes this unlikely in practice: if the police largely determine that certain crimes have occurred by being present to observe them, then the police can only learn from examples of when crime occurs despite their presence.

Common formulations of this problem rest on the unstated belief that police presence will reduce law-breaking, but this is not a given. Crime may occur for many reasons, but differentiating between the following two reasons is crucial in this case: because a population is under-policed and therefore not deterred from breaking the law or the population is simply less responsive to policing. Increasing the risk of arrest through greater police presence would help to address the former, but may do little to help with the latter. Indeed, as Harcourt (2008) notes, there are good reasons to believe that populations exhibit different elasticities with respect to law-breaking in the face of increased police scrutiny, often due to a comparative lack of economic opportunity. Thus, even when attempting to predict crime so as to preempt it, doing so can send the observed crime rate up if additional police presence simply increases the arrest rate.

This formulation of the problem can lead to even further confusion when we consider how to evaluate the efficacy of predictive policing. Do the police want to observe more crime or less crime in the areas to which they have been directed? If the goal is to reduce crime rather than reduce its mere observation, then the model would be most useful when police discover that crime does not occur in their presence. But this would directly contradict the prediction!

This apparent paradox is well captured by former Mayor Rudy Giuliani’s frustration with news in the late 1990s that another city was making a greater number of gun arrests than New York (Reply All 2018). Jack Maple, the inventor of CompStat, the NYPD’s data-driven approach to policing, responded that arrest rates had gone down because crime rates had gone down. Giuliani retorted: “No. Crime goes down, arrests go up.” If we accept that arrest rates are themselves a measure of crime rates, then this statement simply does not make sense. One cannot move in the opposite direction as the other. But if we think of the goal of predictive policing as making predictions about crime so as to increase the likelihood of observing crime, then we might view arrest rates as a measure of police efficacy. In other words: a greater proportion of crimes are leading to arrests. Perhaps Giuliani imagined that this, in turn, must deter crime—that increasing the rate at which crimes lead to arrests must reduce the crime rate. But because the observed crime rate depends on changes to both the rate at which crimes are observed and the rate at which they actually occur, it is not clear what crime statistics actually capture: the police’s ability to detect crime or the police’s ability to deter crime.

Author

• Solon Barocas, Cornell University