\pretitle\posttitle\preauthor\postauthor\predate\postdate

Algorithm versus Judge

I have reviewed several empirical examples of successful and unsuccessful applications of statistical learning algorithms to criminal justice policy. The results of the models considered are associated to gains, such as a potential cut of 2,000 post-arraignment arrests for domestic violence\cite{berk2016forecasting} and losses, such as the use of the Strategic Subject List by police officers to disproportionally target individuals classified as of high risk \cite{Saunders2016}. This spectrum sets the horizon for future applications of machine learning to criminal justice policy questions.
The case of statistical assessments of risk illustrates the complexity of designing a policy instrument to support judges’ decision making process by showing the necessity of algorithmic and policy transparency as it will have an effect over the physical freedom of a fellow individual. The contentious case of modeling an imperfect world at the risk of reproducing its disparities is a hazard of the application of machine learning to criminal justice. However, this hazard is not different from those posed by other methods already used by justice officials to decide, from individual heuristics to clinical assessments, judges already balance competing criteria when deciding over individual sentences and make decision under conditions that vary from case to case and from judge to judge \cite{starr2016odds,laqueur2015machines}. A question to consider is if machine learning or other mathematical approaches to policy might outperform humans and increase the overall efficiency of the system providing a basic framework to support judges -and other public officials- work.
There are several examples suggesting that analyzing the costs and benefits of different policy alternatives, including that of incorporating machine learning algorithms, is safer way to craft policies that yield the highest overall benefits to society, as it is suggested in the economics of crime literature\cite{steve_pato2015role}. Such a framework would allow policy makers to ponder the effect that a particular policy might have in the overall system and, perhaps take action. One example is predictive policing, which has been shown to be effective in forecasting crime hotspots, leading police departments to patrol with higher intensity those areas, which in the process has alerted potential offenders and moved them to other locations, leading researchers to reevaluate the predictive model and make it dynamic, so it can account for potential changes in the overall distribution of crime patterns and police resources deployment \cite{ferguson2016policing,kitchin2015promise}