She predicted murder. Where crime prediction systems have come - and why they are criticized.

[Teacher]

Not so long ago in Chicago, the predictive analytics system believed that Robert McDaniel would soon be involved in a shootout. He was under surveillance: the patrolmen began to visit the shop where he worked, to visit McDaniel at home. In the end, the crime did happen. Robert was assassinated by local bandits, who felt that since the police began to show such attention to him, but still had not arrested him, he was working for them. We do not know for sure if the events took place in this way, because this story became known to journalists solely from the words of McDaniel himself, but predictive systems of analytics have already become firmly established in the life of American law enforcers. Their counterparts in other countries want to follow the example of the United States. Dmitry Serebrennikov, junior researcher at the Institute of Law Enforcement at the European University at St. Petersburg, tells what is known about the structure of these systems - and what objections experts have to justice from the black box.

History
Predictive policing, or pre-policing, collects and analyzes crime data in order to determine a future criminal or predict a crime scene. It is assumed that this allows police officers to prevent possible incidents or to introduce certain preventive measures. In the academic field, such tools began to be developed in the first half of the twentieth century. For example, Ernest Burgess, one of the founders of the Chicago School of Sociology, in 1928 developed a statistical model that predicts the likelihood of relapse after a prisoner's parole. However, such research for a long time did not pass from the pages of scientific journals to the practical plane.
Preventing wrongdoing and even ensuring the "prosperity" of the life of the townspeople was the central task of the European police back in the 17th century, when it first appeared. However, by the twentieth century, especially in English-speaking countries, this principle was transformed into the so-called "reactive strategy of policing", the essence of which is to create conditions for a quick response to violations of the law.
A request for an analysis of the criminal situation and, accordingly, the prediction of crimes appeared with the police only by the 1960s. It found expression in the creation of maps of the concentration of crime and the increasingly close collaboration of law enforcement officers with social scientists (in the case of the United States and England). In the 1990s, this collaboration is dramatically enhanced by the doctrine of New Public Management.
New Public Management is an approach to governance that involves the transfer of models of private corporation management to the public sector, with a focus on data-driven governance. For a more detailed acquaintance with the doctrine as a whole, we suggest referring to this article or, if you are interested in its manifestations in police work, to the texts on the CompStat crime "control" system, which also appeared at that time.

• Modern programs can be divided into two types, both of which are popular with law enforcement officers in the United States and Western Europe: person-based - focused on identifying a person who is likely to commit a crime;
• place-based - focused on the place where a crime is likely to be committed.

The number of “personal” instruments is very large. This can include predicting relapse after parole, analyzing social media and open data in search of potential offenders, calculating bail and punishment for convicts based on their predicted likelihood of re-violating the law, and much more. These systems appeared as a result of the fact that police officers accumulated quite large volumes of various data about citizens - and after the transfer of document flow to digital rails, it became possible to collect and analyze them.
There are many discussions around each of the technologies, which are sometimes only distantly related to each other. However, a common place in them is the opacity of the algorithms, as well as the fact that they may not comply with the laws on personal data and citizens' ideas about the boundaries of personal privacy. Therefore, in some countries (for example, in Germany) the use of such algorithms is limited.
The second type of systems is more versatile. The debate around them focuses on criticizing or supporting the same technical solutions. In addition, the general arguments of supporters and opponents can largely be transferred to person-based solutions, which is why below we will consider only place-based systems.
Their developers, of course, are familiar with the ideas of criminological theories of "broken windows", routine actions and situational crime prevention. However, at the level of practical implementation, police services use models from a completely different science, seismology - which were originally used to predict cascade earthquakes. In this logic, when a sharp increase in certain crimes occurs in some place, one can expect that it will continue for some time, and the police need to have time to respond to it.
Thus, the main task of the police is to control "hot spots" (hotspots) - places with a disproportionately high number and intensity of certain crimes. As one of the first studies on the topic argued, crime is extremely unevenly distributed throughout the city - so much so that 50 percent of calls to the police come from 3 percent of city locations. Through the analysis of the dynamics of such "hot spots", the effectiveness of crime forecasting tools was measured. If the algorithm correctly determined that a hotspot would occur in some place, or vice versa, correctly identified safe areas, then the technology was considered effective.

2014 robbery hotspots in a neighborhood in Pittsburgh, PA

Two giants
The crime scene prediction software itself has evolved from flexible, but difficult to work, to simple, but sometimes clumsy to use, similar to how smartphone interface design has evolved over the past two decades. This is well illustrated by the story of the two giants of this market - the companies PredPol and HunchLab.
HunchLab was created in 2008 and was designed as a multifunctional police assistant - and, as a result, had a not very friendly (but very finely customizable) user interface. As conceived by the inventors, the program was supposed to help the officer analyze different types of crimes and use the widest possible toolkit for this. For example, HunchLab technology provided maps of predicted crimes depending on the chosen statistical model and fine-tuning data for analysis (for example, the current state of crime, weather, socio-economic indicators of different parts of the city, and the like) ...
Sales of the first version of HunchLab failed: out of 60 police departments to which the technology was supposed to be sold, only two were contracted.
The PredPol model was built on a fundamentally different basis - it was created as a system that only needed data on criminal incidents. On their basis, she gave the police a map, on which she highlighted the places where the probability of committing a crime seemed high to her, and he just needed to decide how to go around all the highlighted points as quickly as possible. And this interface fell in love with law enforcement: having entered the market in 2012, the model quickly spread throughout the US police departments.
Curiously, one of the founders of PredPol, Jeffrey Brantingham, is a professor of anthropology at the University of California, Los Angeles (UCLA). He studied the adaptation to the environment of hunter-gatherers in northern Tibet and later, by his own admission, applied the knowledge gained in this field to investigating the criminal areas of Los Angeles - is it so different a hunter-gatherer who sneaks through the forest in search of game? from a "hunter" looking for a badly closed car or a carelessly thrown bag?

Screenshot from the PredPol program

Further development of the systems followed the path of delegating more and more administrative functions to them. After 2012, PredPol and the updated Hunchlab by analogy began to help the patrolmen with the solution of the traveling salesman problem: they built a route and then tracked whether their "wards" had strayed off the route. At the same time, the activities of grassroots police officers began to be evaluated through PredPol, which is where one of the problems of such algorithms arises. It is impossible to create universal criteria for assessing the work of employees, and the management is forced to make a choice. Hunchlab, for example, offers four options for generating metrics:

1. The Police Department independently determines the criteria;
2. The score is calculated from the material damage of the crime discovered by the police officer;
3. The criteria are determined on the basis of the average severity of punishment in court for a particular type of crime;
4. The assessment is formed based on the results of public discussions.

Each of these points, of course, is not perfect. The first gives priority to administrative (and sometimes opaque) logic. The second reduces everything to economic calculation and may neglect the fundamental principles of law. With the third, the opposite situation: focusing on the punishment in court for a particular crime, the assessment strongly depends on the general direction of the state criminal policy or its temporary goals. In addition, the damage from a crime may be disproportionate to the punishment that is imposed for it by the court in principle. The latter principle, in which criteria emerge from public debate, subordinates assessment to highly volatile public opinion. However, similar problems of selection of criteria are inherent in any system of assessment by statistical indicators.
Despite these controversies, PredPol and similar systems quickly gained popularity with law enforcement agencies in the United States and other countries, thanks to the claims of their creators and research that confirmed that the technology reduces the number of crimes. For example, the authors of a randomized controlled experiment conducted in Los Angeles and Kent (UK) concluded that models predict twice as many crimes as criminologists, and their use reduces the incidence of some crimes by as much as 7.4 percent. In this sense, PredPol helped police officers in hot spots, better than the previously practiced “human” analytics of police departments.

PredPol proponents highlight other benefits of using smart tools:

• they are more economical: point prediction allows the police to point and deploy their forces;
• they can mitigate the problem of racial discrimination: initially, the tools were promoted as "color-blind" by definition, and therefore compares favorably with the biased police officer on the street;

Most importantly, the general drop in the crime rate (which such systems are believed to contribute to) increases the sense of security among citizens and, ultimately, improves the quality of life.

Criticism
However, critics of predictive systems show that the picture is not so cloudless. All articles confirming the effectiveness of PredPol were carried out either by its developers or by researchers affiliated with the organization, which gives grounds for comparing the situation with the same opioid epidemic in the United States and the role of pharmaceutical companies in it that sponsored studies on the safety of narcotic drugs. In addition, there are other works, the authors of which do not find a significant effect from new police technologies, but they did not analyze PredPol, but other, less well-known tools. Which brings us to another critical argument.
Because of corporate secrecy, predictive algorithms are already becoming opaque to the police officers who use them. This makes independent verification of the effectiveness of systems like PredPol nearly impossible. Opacity can also play into the hands of police chiefs, making it easier to manipulate statistics and claim a steady decline in crime through the use of algorithms (and thereby legitimizing requests to increase police department budgets).
For example, the Memphis police reported that the system was effective and reduced crime - but later it was found that law enforcement officers were comparing the figures with the year the system was installed, in which there was an unexpected spike in the number of crimes. Re-checking the data in five-year dynamics prior to the analyzed year showed that the real indicators are arranged somewhat more complicated than the police said. Despite the fact that the total number of crimes did decrease by 8 percent - and it is not known what the credit of predictive analytics is - the city's violent crime rose by 14 percent.
Another argument against crime prediction systems is the fear that an algorithm trained on historical data will inherit from the human law enforcement system a certain bias towards certain groups of the population (primarily African Americans in the case of the United States). Such bias can generate causality feedback loop (feedback loop). This is a phenomenon when initially a larger amount of data about a certain area (most often with an increased criminal risk) leads to the assignment of more police officers to patrol it - and this generates a surge in criminal events in the area (because the police need to carry out reporting), which closes circle. MIT professor Gary Marks described the mechanism of such loops as "categorical suspicion" - an officer's deliberately suspicious attitude towards all residents of areas that have been marked by technology as dangerous.
Similar problems exist in person-based systems. For example, in one of the 2018 studies, the authors compared the predictions of the COMPAS system with the responses of users of one of the crowdsourcing platforms to questions about whether a particular prisoner will commit a crime in the future or not. Untrained people had less data than the algorithm, but, firstly, they were able to show results comparable to the program, and secondly, they also showed racial bias.

Accuracy is the percentage of accurately predicted relapses, False positive is the percentage of false positive predictions, False negative is the percentage of false negative predictions (a relapse happened, but it was not predicted). Human - predictions of people, COMPAS - predictions of the program. White bars are "white" suspects. The black bars are African American suspects.

Bias check
One of the most famous demonstrations of algorithmic problems was done by Christian Lam and William Isaac, who simulated police work after the introduction of predictive policing in the event that the initial data for the system is "biased" towards African Americans. For simulations, data on drug arrests in Oakland, California was used: the drug trade there, as shown by qualitative research and polls, is evenly distributed throughout the city, but arrests in connection with it are concentrated in neighborhoods with African American populations.
Scientists tried to replicate PredPol's algorithms and see how they would predict areas where drug-related crimes should occur in the future - and got an overrepresentation of non-white neighborhoods as unreliable, and the algorithm strengthened the police logic, recommending that police visit areas with African American populations twice as often as others.
The creators of PredPol have tried to counter criticism while admitting that the system is not perfect. They first analyzed the results of their own experiment, in which they provided police with data on hot spots once a day, alternately prepared by the program and expert criminologists. The researchers then measured whether the time spent by the police per day in the predicted hotspots was different, as well as the number of arrests in one and the other scenario. In this way, they wanted to show that the algorithm is in no way "worse" than people - and really did not find any significant difference. However, this does not say anything about whether predictive systems project the racial bias of the people they are meant to replace or not.
The developers of PredPol do not deny either the racial bias of the police, or the fact that it can have a significant impact on the operation of technology. One of the co-authors of the program, Jeffrey Brattingham, used simulations to test at what percentage of initially biased data (where bias towards minority representatives manifested itself) predictions would start to change significantly. As a result, he came to the conclusion that the thresholds are 5 percent. The question of what the degree of distortion of real data is still open.
It is difficult to predict how the public consensus will turn in relation to predictive analytics: their popularity among police officers does not decrease, and the developers, naturally, do not abandon attempts to build a more balanced assessment of police actions and the design of predictive technologies itself. Predictive analytics systems themselves are becoming more and more popular outside the United States. In addition to China, which is famous for its experiments with the algorithmization of public administration, such technologies are already being used in the UK, Germany, Poland and many other countries. Not so long ago, the Ministry of Internal Affairs of Russia ordered study of the applicability of machine learning systems to identify the seriality of crimes and offenses. Despite the fact that predictive technologies are in their infancy in our country, we can say about the existence of a general (albeit not so strong) trend for their use by law enforcement officers.
In conclusion, it is worth noting that discussions about predictive analytics in the police often overlook how the police themselves perceive its results and how much it is, in principle, able to influence their decisions. After all, they can, for example, treat the analytics of a multi-million dollar system as a weather forecast - not a very reliable tool that produces non-critical information. But this is a topic for a separate study.

nplus1.ru