What are the pros and cons?
It does sound attractive that based on everything we have learned in the past that is documented in large data sets we can predict what will likely happen in the future. The systems that are currently used have some obvious benefits and Al techniques can automatically distinguish patterns in large data sets. These patterns might indicate that certain locations are more prone to criminal activities than other locations. Or, the data might show that criminal activities have a clear spatio-temporal
Big data and criminology from an Al perspective 3 5 partem meaning that certain criminal activities might happen at certain locations within certain time frames. One can imagine that pickpocketing is more lucrative during rush hour at a railway station than in the middle of the night in a forest. Systems that are interested in predicting spatio-temporal dynamics (i.e., ‘where’ and ‘when’) can be very beneficial for law enforcement. Police officers have a limited amount of time and resources, so a system that allows them to work in a more focused maimer would seem to be desirable.
However, systems that tend to predict who is more likely to commit crimes in the future are more prone to critique. When the police focus their attention on a location where no crime activity occurs no one really gets hurt. But what if the police use a system that predicts that someone is more likely to engage in criminal activity than someone else and the system is wrong? This would have quite serious consequences. If the data sets we base our assessments on are not objectively collected or if there are pre-existing flaws in the data sets, then the system will learn based on faulty information and come up with an algorithm that is possibly biased. If the police mainly focus their attention on a specific minority group (e.g., based on race) in the first place, that group will be overrepresented in the data set. The focus of the police will then determine that the algorithm will indicate race as very important while making assumptions on future crime predictions. However, that does not mean it represents reality, and the minority group may not in fact be more criminally active than other groups.
COMPAS8 (Correctional Offender Management Profiling for Alternative Sanctions), developed by Northpointe Inc., is an example of a machine-learning model that has been critiqued for producing biased predictions.9 It is a risk-assessment algorithm used to predict hot spots of violent crimes, determine the types of supervision inmates might need, and provide information that might be useful in sentencing. COMPAS is also referred to as a black box algorithm, meaning that the developers do not reveal the details of their algorithm. In effect, it is impossible to reliably assess the COMPAS system when the details of the algorithm are unknown (Skeem & Louden, 2007). In short, when the developers chose to keep the details of the algorithm to themselves, the key question becomes: how can outsiders trust the system? They developed a system that may have a big impact on the life course of people, but they refuse to divulge any insight into their methods of prediction. Making predictions of future crimes is already a delicate topic, but if the decision-making algorithm is based on a black box it will likely trigger even more discussion. An investigation by ProPublica10 found evidence that the COMPAS model might be biased against minorities, but this cannot be confirmed or rejected until full details of the algorithm become public.
Trap Wire
Another methodology concerned with detecting thr eats and prevention of terrorist and criminal events is known as Trap Wire.11 In the aftermath of the attacks of September 11th in the United States, a key government priority was to develop a method that would prevent such (terrorist) attacks from happening in the future. The
Trap Wire software uses algorithms and data from a number of surveillance sources to predict potentially criminal/tenorist behaviour. Trap Wire recently gained negative attention through the WikiLeaks files that claimed that the software applied a methodology based on a much more widespread and extensive breech of privacy than people were aware of.1213 This illustrates another potentially negative aspect of the use of big data in crime prevention: what about privacy?
In order to build a ML model, extremely large data sets need to be collected. This means that many things you do might be monitored, for example, where you go, what you buy, and who you are in contact with, to name only very few. All this data is in turn gathered because there might be an evident association with criminal activity, but in many cases, there probably is no connection whatsoever. Nevertheless, all of the data are requir ed in order to reach these conclusions. But are we willing to hand in a large part of our privacy for a black box method that might incriminate ourselves? To what extent is that even feasible?
