Several German federal states are testing software for predicting crimes, others are already using it. The risk of „danger“ or the recidivism of offenders is also to be calculated. However, a reduction in crime with the help of computer forecasts cannot yet be proven reliably. Instead, the applications are loaded with prejudice.
Predictive policing is an attempt to calculate the probability of future crimes based on near-repeat theory or the assumption of repeat victimisation. Similar to the „Broken Windows“ theory, it is assumed that earlier delinquent actions are likely to be followed by others. Data on crime scene and time, prey and procedure are processed and weighted according to a certain procedure (scoring). Data mining is used to identify patterns and find serial offenders.
According to this logic, the limit of predictability is not determined by the algorithms, but by the computing power of the computers or the data sources that are included in the analysis. In fact, a study commissioned by the State Criminal Police Office of Lower Saxony points out that predictive policing is ultimately a further development of crime mapping, which police authorities used to use to digitize their pins on the map. Only the weak computers would have prevented the data from being correlated.
The federal states of Bavaria, Berlin, Baden-Württemberg, Hesse, Lower Saxony and North Rhine-Westphalia have already introduced applications for predictive policing. Lower Saxony, Brandenburg and Hamburg are initially conducting studies. The Federal Criminal Police Office coordinates the efforts in a federal-state project group and initiates research on a European level. All the projects start with the phenomenon of burglary: Here the readiness to denounce by the citizens is high, and a high percentage of the crimes might be repeat offences. The respective modus operandi could allow conclusions to be drawn about expected events in a supraregional comparison.
„We must get before the situation“
„What is so critical about the police using such instruments in a digitised society,“ asks Dieter Schürmann, director of criminal investigation for North Rhine-Westphalia Predictive policing follows the shift in the fight against crime to the forefront, as the former BKA president Ziercke formulated it more than ten years ago with his motto „we have to get before the situation“. With the use of more and more digitally generated and stored data, the police in the digital field often act hard at the limits of what is permitted. The automation of police danger defence through the introduction of forecasting software is likely to reinforce this trend. More than two years ago, the Conference of Federal and State Data Protection Commissioners warned against a „further shift of the police intervention threshold into the forefront of dangers and crimes“. Today it is completely unclear which crimes will be detected automatically in the future and which data sources will be included. In digital investigations, the haystack must be enlarged in order to find the needle. There is also the danger of erroneous forecasts, which, according to the data protection officers, is to be expected particularly in the case of the increasing apron analysis and is associated with significant effects on the suspected persons.
In Bavaria and North Rhine-Westphalia, the purpose of the software is to be extended to other crimes in the public arena, with car theft or robbery being discussed. The data sources will also be expanded. Currently the weather, traffic data or expected events can be processed. From a police point of view, however, not only these data are relevant. Meaningful are, for example, the connection of an area to motorways or local traffic, or information on buildings.
The criminal investigation departments also want socio-economic data on income distribution, purchasing power and creditworthiness or the value of buildings to be used. Some authorities have already been supplied with this information by statistical companies. Current water and electricity consumption can also allow conclusions to be drawn about criminal offences, as this indicates the absence of the occupants. The Institut für musterbasierte Prognosetechnik (Institute for Pattern-Based Forecasting Technology), a company from the German city Oberhausen, is testing in Baden-Württemberg whether their „Precobs“ software can be improved with information on the proportion of foreigners in a residential area.
Broad spectrum of data sources
The North Rhine-Westphalian state criminal director is thinking aloud about the use of „real-time data“ from telephones, although it remains unclear whether this means the serial numbers of the SIM cards used or whether Schürmann would like the surveillance of mobile phones to be included into the prediction by telephones that have been detected near a crime scene. It is also conceivable that license plate recognition could be integrated, as is already the case in some US cities. The software company Microsoft offers a „Domain Awareness System“ for this purpose, which is intended to find suspicious persons as well as vehicles searched for and process information from other sensors, including cameras, for example. In this country, such an application would rather operate as a pimped-up police control centre, but Microsoft markets the system as predictive policing.
Finally, openly available information from social networks can also be integrated. The police authorities themselves could supply such pre-filtered data. A modern police control centre today has functions for evaluating trends on Twitter, Facebook or Instagram. Such software comes from the Oberhausen-based company rola Security, which has since been taken over by T-Systems. The police could use it to track hashtags or geodata on Twitter during a mission. For example, a georeferenced display of tweets from football fans or demonstrators would be advantageous for the assessment of the situation in order to draw conclusions as to what action will soon be necessary. The manufacturer of „Precobs“, is developing a system against „graffiti sprayer gangs“ in cooperation with the Swiss company Futurelab in Zurich, which evaluates Facebook, among other things.
On the other hand, the results of predictive policing also end up in social media. The Institute for Pattern-Based Prediction Technology has developed an Android app for Precobs, which is used by the Swiss canton of Aargau under the name „KAPO“ („Kantonspolizei“). Under the motto „The police warns“, its users can be informed with push messages about alleged impending criminal offences in their own neighbourhood. The reports about crimes that have not yet happened make the population a block warden.
Find out who is radicalizing themselves?
So far, predictive policing in Germany has only used statistical information. In countries like the USA or Great Britain, however, the software has long been used on a personal basis. For example, personal data is processed in order to determine so-called „heat lists“. In the case of emergency calls, some cities in the USA calculate the danger of their engagement by comparing the calling person with the criminal record. Persons on a „Strategic Subjects List“ of the police in Chicago are considered particularly at risk of being involved in a shooting. As persons at risk, they are to be visited more frequently by the police. Social programs are also intended to catch the people at risk. The RAND Corporation criticizes that the potential perpetrators on the list have a higher risk of being arrested instead of being supported with social support programs.
In the US, judicial authorities in some states calculate the likelihood that a offender will relapse. The result can influence the period of detention. The British police also tried to calculate the recidivism of offenders. Accenture, the company commissioned to do this, presented its study at the „European Police Congress“ sales fair, after which representatives of German state criminal investigation offices took part in a confidential workshop.
At the beginning of the year, the BKA launched a project to calculate the dangerousness of potential criminals. The application RADAR-iTE developed for this purpose processes data of so-called „perpetrators of danger“ and is intended to calculate which might be planning an attack and should therefore wear a shackle. In the process, data from social networks could also be used to find connections between people and to deduce who is currently radicalising themselves. However, RADAR-iTE is based on Microsoft Word and Excel. So this is probably not an algorithmic application. Rather, the information is evaluated according to a point system and then flows into an overall assessment.
In further research projects involving the BKA, possibilities are being explored for searching the Internet and social media for „radical propaganda and propaganda“ With „indicators for the early recognition of radical tendencies“, the criminal investigation offices, with the support of software companies, want to recognise and combat the development of „radicalisation processes“ as early as possible. In one of the projects, an „analysis and evaluation instrument“ and software for „recognition of extremist network structures“ are to be developed.
Focus on people from Southern and Eastern Europe
The data processed in predictive policing come from police crime statistics, which can be tendentious. The data counts complaints, not actual crimes. If, on the basis of these data, the police check people with a certain appearance or at social hotspots more frequently, more crime reports are also recorded there. These are included as case statistics in the prediction of crimes and confirm the apparent assumption that crime is on the increase in these neighbourhoods or by these groups of people. The U.S. manufacturer of „PredPol“ even grants a discount if the police disseminates the evaluations carried out by the company (on the basis of police statistics) in press releases.
The software used for predictive policing does not say what burglars who are to be held by the police actually look like. So the police lie in wait after the „usual suspects“, which cements already existing classisms and racisms. A television report by the ARD studio Washington had impressively demonstrated this for the introduction of the software „PredPol“ in Santa Cruz. In the (unfortunately no longer available online) article entitled „‚Minority Report‘ wird Wirklichkeit“ (Minority Report Becomes Reality) only „conspicuous“ people were cut into it, such as people wearing hoodies, neglected clothes or dark skin colour. Also the investigative journalist office Propublica proved that persons with dark skin color are indeed systematically disadvantaged by the (mostly unknown) algorithm.
Similar stereotypes can also be observed in Germany. In the German federal states, the state criminal investigation offices want to pester the target group of „travelling serial burglars“. Which is not said: In the international context, the term stands for „Mobile Organised Crime Groups“, which usually refer to so-called „perpetrators“ from Romania and Bulgaria. German predictive policing thus focuses on people whose origin is suspected primarily in Southern and Eastern Europe.
Functioning of the algorithms is unknown
Evidence that predictive policing leads to a reduction in crime in a certain area has not been available so far. There is a lack of robust research. This is also pointed out by the Lower Saxony police, who commissioned the aforementioned study. So far, only perceived effects can be determined. Two investigations are to bring light into the darkness: A „study of new technologies for predicting criminal offences and their consequences for police practice“ is currently being developed at the University of Hamburg, but the project will not end until December 2018. The evaluation of a predictive policing project in Baden-Württemberg by the Max Planck Institute for Foreign and International Criminal Law in Freiburg has now been completed.
According to the study, it is still „difficult to judge“ whether the tested software PRECOBS can contribute to a reduction of home burglaries and to a trend reversal in trap development. The „crime-reducing effects“ would only be in a moderate range and could not be significantly reduced by predictive policying alone. In some areas of the pilot area the number of burglaries decreased, in others there were increases. Most predictions also concerned urban areas with higher burglary rates and thus more statistical data. The evaluation study therefore takes a critical view of the benefits for rural areas.
Control often fails
Public or parliamentary control of digital security regularly reaches its limits. In most cases, no statistics are kept on the use of software. Data protectors also initially only check whether personal data is being processed and if so, whether it is being used for the correct purpose. However, the selection and combination of the incoming information could have a discriminatory effect – even if no personal data is used. In addition, the data could be taken out of context. The self-learning capability of modern software, commonly referred to as artificial intelligence, is likely to make this a very dangerous situation. The presumption of innocence is then replaced by machine logic, as the Hamburg data protection commissioner Johannes Caspar puts it.
Another problem is that the manufacturers do not disclose the source code of their software. It is not possible to check how the algorithms actually calculate and weight their forecasts. Those affected cannot defend themselves against a possibly falsifying classification. The data protection commissioners of the Federal Government and the Länder therefore rightly point out that the constantly further developed technical evaluation possibilities „already today hold the potential for citizens to lose control over their data – to an extent and in a way that was inconceivable in the past“. A political debate on predictive policing is therefore needed. Once introduced, the calculation of burglaries or „endangers“ can gradually be expanded into an instrument of social control.