Page 14 - Case study: Crime prediction for more agile policing in cities – Rio de Janeiro, Brazil

Predictive policing is grounded on several established theories of crime behavior and crime
opportunity explaining crime concentration and repetition, and why crime occurs in some places
and not in others. These include:

• Routine Activity Theory that states that crime depends on multiple factors including the
motivation of offenders, suitable targets and an absence of capable guardians ;
12

• Rational Choice Theory underlines that criminals make rational decisions based on opportunity
and estimated costs such as the possibility of being imprisoned and punished ; and
13
• Crime Pattern Theory that explains why, when and where crime happens, focusing on the
intersections and commonalities between victims and perpetrators .
14
By applying different combinations of these methods and theories, a growing number of universities
and commercial vendors have developed predictive policing software, serving to some of the
world's largest police departments. Existing commercial solutions can be broadly categorized into
two categories: (i) methods that predict the location of crime; and (ii) methods that predict likely
crime offenders and victims.

The first method involves processing historical police reports, emergency hotline calls, weather
forecasts and even the locations and dates of large public events to calculate the probability of
crime happening in space and time. The second - and perhaps more controversial - method often
processes arrest data including criminal records and social media profiles as well as location, race,
15
age, gender, and ethnic data. The method generates a shortlist of high risk individuals who are
determined to be potentially involved in future crime.

Controversies associated with crime prediction: There are widespread concerns that predictive
policing tools could unintentionally exacerbate over-policing of marginal areas and undermine
privacy. It is widely known that algorithms can reproduce existing patterns of discrimination,
reinforcing previous errors and biases of programmers and embedded in databases. There
are very real ethical questions about the extent to which such tools can influence police to
disproportionately surveil marginalized neighborhoods and communities. Related, there are fears
that such tools may augment race and age profiling and undermine privacy rights and civil liberties.

Recent studies funded by the US National Science Foundation demonstrate how predictive policing
models are susceptible to 'runaway feedback loops' . In these cases, police are repeatedly sent to
16
the same identified hot spots, irrespective of the true crime rates. The researchers demonstrated
how historical crime incidents that have been "discovered" by on-duty officers can aggravate the
degree of runaway feedback, while in turn, historical incidents that were "reported" by citizens can
attenuate, but cannot entirely remove such feedback.

The accelerated pace and spread of crime and violence prediction tools means that these concerns
will only grow in the coming years. Indeed, new platforms are already being tested that aim to
automatically classify gang-related crime , combine social media with criminal history to predict
17
crime , and use artificial intelligence to identify individuals with higher risk profiles of committing
18
19

8 Crime prediction for more agile policing in cities – Rio de Janeiro, Brazil