You are on page 1of 7

Predictive Policing

Dheeraja Rajreddygari | 2020101120


Sanika Damle | 2021115005

| Introduction
With advancements in artificial intelligence and machine learning, a lot of traditionally manual
jobs have been replaced with technology. Although the technology we have chosen does not replace jobs,
it attempts to maximise the potential of the police forces. While looking for technologies for this essay,
we stumbled upon the movie “Minority Report”, which is about trying to predict and prevent future
crimes. Although this movie heavily relies on supernatural abilities, predictive algorithms that work on
that idea have been created and there have been attempts of implementing the same in a few US cities,
some countries in Europe, and India. The algorithm of predictive policing is an attempt to prevent future
crimes by mathematically analysing large sets of data, to help decide which individual may be a
“high-risk” person, or to decide where to deploy more police. It uses historical as well as current data to
try and predict which people are more likely to be victims or perpetrators of future crimes. The roots of its
origin are in China, with the policy of social governance which was used to spy on its citizens (Sprick,
Daniel, 2020). The earliest recorded implementation is by the LAPD, called LASER which is a policy
used to predict gun violence (Lau, 2020). Another policy, named PredPol, predicts regions which are
more likely to have crime and labels them as “crime hotspots”. The policy in India is called CMAPS and
while not very widespread, it has started being used in a few states. There are a lot of controversies
surrounding predictive policing, and there is a need to study the algorithm and the ethics behind it. We
feel that if “advancements” in technology are meant to help more people, this piece of technology can be
classified as primitive. In this essay, we aim to address some of the issues around predictive policing and
analyse its impact on society.

| Impact on Society
The main goal of predictive policing is to be able to redirect police forces to be able to use them to their
full potential. If implemented correctly, in countries like India, which has one of the most understaffed
law enforcement in the world, it could be a welcome change. However, the impact of this algorithm on
society has mostly been negative. One of the major issues that arise is that law enforcement loses
accountability in the way that they patrol places. If they tend to monitor places where, say, black people
live more closely due to personal racial prejudices, they can blame the software for directing them to that
place. This can be very problematic in countries like the USA where minority communities already have a
history of rocky interactions with the police.

LASER is supposed to be a representation of a laser surgery, which removes tumours, which in this case
are perpetrators. However controversial this implication is, it does not compare to the actual working of
the program. It was primarily made to try and prevent gun violence and used an algorithm to detect places
where crimes have frequently occurred and also makes a list of people that it considers are more likely to
commit crimes (The Associated Press, 2014). For identifying places, the algorithm uses data on past
crimes that have occurred at that location.

The list of people is based on scores that are assigned to each individual based on past crimes, gang
memberships and interactions with the police. The problem is that “interactions with the police” can even
be them approaching them for help. What this does is it creates a sense of distrust between the two parties
and people do not freely approach the police, even when they need help. This system also makes it
impossible to get off the list once you are on it. Because the area is targeted more by law enforcement if it
is considered “high risk”, it leads to minor infractions with the police leading to more points being added.

PredPol, now known as Geolitica, focuses on finding crime hotspots. It uses three parameters, crime type,
crime location and crime data/time which are fed into the algorithm. (PredPol, 2021) It focuses on
identifying regions where crime is likely to be committed, rather than the individuals who commit crime.
It claims to be free of racial biases since regions have no notion of race. This algorithm also allows the
police to justify and perpetuate the cycle of systemic bias. The bias against minority communities is now
seen as a result of the algorithm, which is perceived as objective, rather than subjective human judgement.

While some police claim that this is a “huge step forward”, and that they would not be able to do their job
without this technology, their point of view is extremely one-sided (The Associated Press, 2014). Just like
any piece of technology, people that benefit from it will be in support of it. If the police succeed in
reducing crime in an area, the residents will welcome it. People that are negatively affected by it are
mostly minority groups, and due to past police experiences, they avoid speaking out.

This system creates a lack of “digital trust” as individuals cannot be sure of the data that is being used in
the algorithm. There are major issues with the ethics of the implementation as they use personal data, but
in most cases, people are not made aware of this. Since the algorithm can also use other factors like
CCTV footage, residents need to be made aware of the fact that they are under surveillance. In most
cases, they are not aware of this breach of privacy, and cannot consent to it. There are a lot of ethics with
the implementation because of factors like this, and if governments are planning to implement these on
larger scales, new laws need to be put in place, as otherwise, it is a violation of fundamental rights.
Another question is the problem of anonymity. Algorithms like PredPol that use data for certain areas
claim to keep the identities of the individuals that they flag as anonymous. However, we do not know if
this is always followed, as the process used is not transparent.
| Biases
The system of predictive policing currently in use is biased on two levels: there is bias in the model itself,
and then there is the bias introduced by how the data is used.

All machine learning models are vulnerable to biases introduced due to the training data used. A common
phrase used to explain this phenomenon is “Bias in, bias out”, meaning that if the data fed into a model is
biased, the predictions made by it will also exhibit those biases. The bias in the data used by PredPol is
two-fold. First, the available data on crime used to train the algorithms includes only those crimes
recorded by the police. Unfortunately, there is a significant gap between crime that is recorded and crime
that actually occurs. A survey by the US Bureau of Justice Statistics found that victims of crime tend to be
apprehensive about reporting the incidents to the police. The survey estimates that only 40% of violent
victimizations and 33% of property victimizations were reported to police in 2020. (Bureau of Justice
Statistics, 2021) There is also a demographic difference in the proportion of crime reported; white victims
of violent or property crimes were less likely to report the incidents to the police. This difference due to
under-reporting of crime is made worse by the fact that not all crime reported to the police gets recorded,
due to the perceived insignificance of the incident. This gap between recorded and perpetrated crime
makes police data an unreliable basis for predicting future crime. This disparity leads to over-policing in
regions where crime is well-reported and well-documented, and under-policing in others. Second, there is
a historic bias against African Americans and Latinos. In many cases, they are falsely convicted and
imprisoned without a chance at a fair trial, or receive harsh sentences for minor infractions. Since the
models use this data to predict trends in crime, they are biased to predict that these groups are more likely
to commit crimes.

This inaccuracy in data makes predictive policing algorithms fundamentally flawed, in the sense that they
do not accurately understand the problem that they are trying to solve. The algorithm used by PredPol is
based on a machine-learning model that was designed to predict earthquake aftershocks (Mehrotra et al.,
2019). However, there is a fundamental problem here. Earthquake data is measured by seismographs
located all over the world, meaning that every earthquake that happens is recorded and accounted for. The
model works because we have exhaustive data. Since the same cannot be said about police records of
crime, this model is not a good fit for predicting crime. Any attempts at predicting future crime based on
past records must take into account any flaws, biases and gaps in past data.

Apart from the flaw in the model, there is further bias introduced due to the way the system is used: a
region or individual that was identified as being “high risk” would be monitored incessantly, without a
chance of redemption. PredPol identified certain regions as hotspots of crime and deployed more police
patrols to those regions. However, this meant that more crime would now be reported and recorded in
those regions, just because the police happened to be there when the crime occurred. So crime that would
have gone unnoticed otherwise was now being recorded in these “hotspots”, while nothing changed in the
other regions. Since more crime was being recorded from these hotspots, they continued to be labelled
that way, resulting in a continuous feedback loop of over-policing. With LASER, individuals labelled as
“chronic offenders” based on their history are targeted and monitored by the police to prevent them from
actually committing the crimes. Since the police tend to intervene and conduct frequent checks with
chronic offenders, these interactions with the police were recorded in the system. This made them more
likely to receive more points, making it almost impossible for individuals to get rid of the label once they
were identified as chronic offenders.

| Controversies
The model used by predictive policing systems is questionable and raised concern among mathematicians
and other experts in the field of machine learning. In a letter that was published in the October 2020 issue
of Notices of the American Mathematical Society, some mathematicians called for a boycott of
collaborations with the US police department (Linder, 2021). They explain that the models and algorithms
used to predict crime do not accurately predict real-world data, claiming that it simply perpetuated the
existing biases and calling it a “scientific veneer for racism”.

While police departments claim to be benefitting from the use of predictive policing, this is not an opinion
shared by experts. The prediction algorithm used by PredPol, despite looking complex, essentially takes
the “average” of the locations where crime has occurred and tells the police to patrol those areas
(Mehrotra et al., 2019); it basically tells the police to keep going back to the regions where they already
found crime. This does not add much, since the fact that we have records of crime from a region means
that the police are already patrolling it. The predictions it makes are only telling the police to patrol areas
where the kind of crime they already catch occurs.

While we do not know the full extent of data used by LASER, reports indicate that officers are required to
fill out field cards, which are fed into the system (Bhuiyan, 2021). This means that almost any interaction
with a police officer–including random stops, reporting, witnessing or being a suspect in a crime without
being convicted–increases the risk of an individual being marked as high risk. This has the potential to
make the public fear the police and make them reluctant to approach them. This is doubly dangerous for
minority communities and financially backward groups who are already more likely to be victims of
crime and at a higher risk of being falsely convicted since they are now unable to report any crime
perpetrated against them without risk.

Another cause for concern is that the system is not transparent. In the studies conducted by Gizmodo and
Vice, the police departments of many states were reluctant to share the details and statistics related to their
predictive policing systems. Criminal defence lawyers are not informed when an arrest was made due to a
crime prediction system (Aaron Sankin, 2021). Giving the police the power to monitor and target
individuals and regions without having to explain themselves can push the already strained public-police
relationships in the US.

An audit by the Inspector General of the US Police Commission found many alarming insights into the
operation of predictive policing (Los Angeles Police Department, 2019). The audit found that from state
to state, there were inconsistencies and lack of proper documentation about the specific kind of data being
collected, and the status of each individual. Individuals identified as chronic offenders could be marked
active or inactive by the officers at any time, and no record was maintained about when the changes were
being made and by who. This makes the system open to human subjectivity and reduces the credibility of
the department’s claim about relying on data and algorithms. The audit also found that contrary to what
was claimed by the states’ police departments, some regions were not assigning points at all, and were
selecting chronic offenders based on referrals from detectives and patrol officers. As a result, 112
individuals with zero points were added to the list of chronic offenders. This is an extremely concerning
finding which completely lays waste to the department’s claim of objectivity and lack of racial bias due to
the use of algorithms.

There are numerous cases where the lack of transparency and biases in the prediction systems have led to
unjustifiable over-policing and harassment by the police. We explored one case study of Buena Vista,
which is a housing complex in Elgin, Illinois consisting mostly of low-income African American families.
In the 1.5 year period from January 2018 to October 2020, the police made 121 arrests at the complex.
This includes arrests made for domestic abuse and several minor infractions including trespassing. Based
on this data, PredPol’s software predicted that an astounding rate of 100 crimes a month would occur here
over a 29-month period. This is made worse by the fact that the algorithm predicted only 154 crimes for
the same 29-month period in a nearby region where White residents are the majority.

| Future Implications
Apart from all the biases and controversies mentioned, there are a lot of flaws in the model which can be
exploited. Data leaks can be devastating and data tampering can also result in people being wrongfully
convicted. However, there are ways to make the system better as well.

If algorithms like these are to be implemented on a larger scale, care must be taken to deliberately remove
biases from the data. Countries like the USA have had protests for a change in the legal system. If the
model is trained on this racially prejudiced, the predictions it makes will also represent that. If the legal
system has to be improved, the model cannot be trained on past data, so the data passed to it must be
carefully crafted to remove this bias. The functioning of the models depends purely on the data provided
to it.

Another thing that should be reinforced is redefining what “interactions” with the police entail. The
system should not target people who have been convicted in the past and have served their sentence, as
people are capable of change. The current policy inadvertently targets people with minor convictions and
denies them the chance for redemption.

For the algorithm to be improved, it is also important to identify and remove short-lived trends in crime
data that are no longer relevant in the longer run. Certain social, political and economic conditions can
affect the nature and extent of crime for a short period, however, they do not reflect long-lived trends in
crime and must be identified and accounted for by the algorithm. For example, elections and election
campaigns may result in more gang-related crime as opposing gangs fight for political control over a
region. But this short-lived increase in crime may not be sustained post-election.
Another change that should be made is a change in the laws surrounding predictive policing, to make law
enforcement more accountable, and the process should be made more transparent. This can be done
through regular auditing, and the results of the audit must be made available to the public.

In Europe, in 2022, there was a call to ban predictive policing as they argued that the system
disproportionately targeted the most marginalised people in society, infringed fundamental rights and
reinforced structural discrimination (Skelton, 2022). If people are not successful in removing biases,
banning predictive policing seems like the most rational idea.

| Conclusion
Currently, the algorithm has a lot of flaws, and for it to be used in the future, there are a significant
amount of changes that need to be made to try and make it fair and reliable. There are decades of systemic
bias that make up the dataset that is used to “predict” future crimes. This bias has been shown to target
minorities. Increasing police patrolling is not a failsafe way to prevent crime, as there is a noticeable gap
between the crimes that are predicted and the crimes that are reported. More focus on an area will just
lead to people being penalised for minor interactions, and the score of that place will increase, which will
result in more police in the area. It is a never-ending cycle that benefits no one. To improve the model,
there are other factors that can be taken into consideration, but even then there is no concrete proof that it
will work. Transparency and accountability by the law enforcement is the bare minimum that they can do
and making people aware of the system of predictive policing is also essential.

Even after all the changes, it, at its core, will be a model that uses past crime to predict future crime.
There is no proof to validate the working of the model, so it has no real scope even with improvements.
Predicting crime may be possible but using historical data to do it is not the way to go, as there have been
mathematical papers disproving this theory. This can be considered an example of a technological
advancement that has actually been a step back in progress.
Citations
1. Aaron Sankin, D. M. (2021, December 2). Crime prediction software promised to be bias-free.
new data shows it perpetuates it. Gizmodo. Retrieved May 6, 2023, from
https://gizmodo.com/crime-prediction-software-promised-to-be-free-of-biases-1848138977
2. Bhuiyan, J. (2021, November 8). LAPD ended predictive policing programs amid public outcry. A
new effort shares many of their flaws. The Guardian. Retrieved May 6, 2023, from
https://www.theguardian.com/us-news/2021/nov/07/lapd-predictive-policing-surveillance-reform
3. Bureau of Justice Statistics. (2021). Crime in the United States, 2020: Uniform Crime Reports.
Retrieved from https://bjs.ojp.gov/sites/g/files/xyckuh236/files/media/document/cv20.pdf
4. Lau, T. (2020, April 1). Predictive policing explained. Brennan Center for Justice. Retrieved April
20, 2023, from
https://www.brennancenter.org/our-work/research-reports/predictive-policing-explained
5. Linder, C. (2021, November 2). Why hundreds of mathematicians are boycotting predictive
policing. Popular Mechanics. Retrieved May 6, 2023, from
https://www.popularmechanics.com/science/math/a32957375/mathematicians-boycott-predictive-
policing/
6. Los Angeles Police Department. (2019). Board of Police Commissioners Report: BPC #19-0072.
Retrieved from https://www.lapdpolicecom.lacity.org/031219/BPC_19-0072.pdf
7. Mehrotra, D., Mattu, S., Gilbertson, A., & Sankin, A. (2019, February 14). Academics confirm
major predictive policing algorithm is fundamentally flawed. VICE. Retrieved May 6, 2023, from
https://www.vice.com/en/article/xwbag4/academics-confirm-major-predictive-policing-algorithm-
is-fundamentally-flawed
8. PredPol. (2021, March 2). Predict prevent crime: Predictive policing software. PredPol. Retrieved
May 6, 2023, from https://www.predpol.com/
9. Skelton, S. K. (2022, April 22). EU lawmakers propose limited ban on predictive policing
systems. ComputerWeekly.com. Retrieved May 6, 2023, from
https://www.computerweekly.com/news/252516238/EU-lawmakers-propose-limited-ban-on-predi
ctive-policing-systems#:~:text=On%201%20March%202022%2C%20Fair,rights%20and%20rein
force%20structural%20discrimination.
10. Sprick, Daniel (2020-09-22). "Predictive Policing in China". Naveiñ Reet: Nordic Journal of Law
and Social Research (9): 299–324. doi:10.7146/nnjlsr.v1i9.122164. ISSN 2246-7807. S2CID
254483644.
11. The Associated Press. (2014, November 14). LAPD uses big data to target criminals. CBS News.
Retrieved May 6, 2023, from
https://www.cbsnews.com/news/lapd-uses-big-data-to-target-criminals/

You might also like