These developments herald the potential chaos biased government systems can cause by the twisted use of modern analytical methods not limited to any particular nation but something which can affect law enforcement around the world.
To elaborate the term in our title ‘prediction policing’ is the use of mathematical analytics by law enforcement to identify and deter potential criminal activity. [2]
Building on the above statistical or machine learning models of analysis based on the likes of the lasso, elastic net, logistic multinomial regression, or in some cases support vector regression being fed with police records of time, location, and nature of past crimes may seem to provide a solid foundation for such predictive policing based on which the law enforcement can decide when, where, and who may commit the next crime and spot the potential crime-prone neighborhoods to regularly patrol, but human actions and social dynamics can only be assessed after a particular attempt once made or expressed upon and cannot be left in the hands of a mathematical model. One can imagine the consequences if an AI system is built around the same. It will lead to what we call a ‘problem of high bias’. This shall not be surprising given the cases of gender bias during recruiting by AI systems that recently surfaced and in this context - racial or other officer biases.
Keeping the above concerns in mind the aforementioned letter proposes a few guidelines for a chance of better predictive policing (which again can be considered by the police around the globe):
Any "potential high impact" algorithm to be put to a public audit.
The audit process should involve experts for active use of mathematics to "prevent abuses of power."
Mathematicians should work with community groups, oversight boards, and other organizations involving minorities dealing with AI and data science to develop alternatives to "oppressive and racist" practices.
Learning outcomes addressing the "ethical, legal, and social implications" of such tools should be implemented in the data science courses offered by the academic departments.
This is not to say that such an approach should not be applied at all and as is prevalent in several police departments in several places tools of predictive policing can come in handy if the data based on heuristics related to reported crimes and under-watched or under-patrolled areas are mapped well, correlated with areas of previous criminal incidents and then an algorithm is developed to disperse the patrol units accordingly on those predicted crucial locations. Also to be noted is the general public mentality of not reporting crimes because either police is considered to only cause more trouble or there is not enough cooperation by the departments to lodge a complaint around the world which must also be addressed for the police departments to gather more relevant data for the functioning of the algorithm. The figures below give a demo of how a current biased tool would map the crime hotspots in contrast to how it would look like if the suggested model is implemented.
Fig. (a) Potential crime hotspots mapped on the basis of one of the predictive crime algorithms in question.
Fig. (b) Potential crime hotspots mapped on the basis of the suggested algorithmic model based on the data type mentioned above.
References