A red square lights up on the display supplied with the police car, indicating in which area of the city and in what time slot a crime is most likely to be committed. A statistical forecast made by analyzing thousands of data – where the most recent crimes were committed, when, in what circumstances, in what weather conditions and more – and putting machine learning algorithms to work.
US PredPol, HunchLab, Palantir, German Precobs and many others operate – in a nutshell – in this way: using so-called hotspot analysis to report to the police the hot areas of the city. A mechanism that has raised several concerns about its actual usefulness: in fact, if more crimes are reported in an area – and consequently more police are sent – inevitably more crimes will be identified; consequently making that same area even more subject to control.
A vicious circle of predictive police – as pointed out by The Verge – is likely to be ineffective, to leave completely uncovered other areas of the city and put under more pressure certain communities, reinforcing fears that these algorithms can exacerbate prejudices already present in society.
The Italian exception: KeyCrime
All predictive police software works this way. All but one: the Italian KeyCrime, conceived by the former assistant chief of the Milan Police Headquarters Mario Venturi, who – after having worked for decades in the police – has embarked on a new career as an entrepreneur, transforming his software into a startup. “I started working on it in 2004. Four years later the experimentation began, when the then commissioner decided to use my crime analysis system to fight commercial robberies,” Venturi tells Wired.
2008 was indeed a challenging year for retailers. Only in the city of Milan there were 664 robberies (excluding banks) “It is a type of crime that has a tremendous impact on citizenship, whose greatest damage is not so much in crime against property, but the one against the person,” says the former policeman. “I saw owners close their businesses because, after suffering a robbery, they lived in terror. Since this is a particularly widespread and hateful crime, it was decided to experiment with software in that area. In the first year of testing we have identified the responsibilities of 47% of the robberies. ”
It is a figure that continues to grow over the years and has now reached 60%. “These are numbers that are unmatched worldwide,” says Venturi, and which have guaranteed KeyCrime partnerships with companies such as IBM, scientific validations by institutions such as Essex University or the National Bureau of Economic Research in Boston and have grown the ambition to spread this software – which is used at the moment, in the province of Milan only , and of which a new version will be ready very soon – globally All this, also relying on 1.2 million euros of investments provided by venture capital company Oltre, the investor shareholder Sdg Group and the entrepreneur Giorgio Gandini.
The role of crime linking
But what is it that differentiates KeyCrime from other predictive police software? “The idea came by looking, for work, a mountain of files relating to various crimes, in which data were collected badly but still contained information that would allow us to hypothesize behind such crimes, although occurring at different times and places, there were the same hand, ”says the founder of KeyCrime.
Behind hundreds of robberies, in fact, there are not hundreds of robbers, but a few dozen. Instead of using the hotspot analysis to predict in which area of the city crimes could occur (with all the limits already highlighted), it would not be better to use software to understand which crimes are committed by the same robbers and predict where, how and when these robbers will perform their next actions?
This is the fundamental feature of KeyCrime, which analyzes thousands of data elements (eg: where the robberies took place, at what time, how the robbers behaved, what vehicles and weapons they used, how they were dressed and much more) to identify correlations between different crimes and determine which ones were committed by the same person or group of people. Crime linking, therefore, is the fundamental aspect of the software designed by Mario Venturi; it allows users to find the crimes carried out by the same subjects, connect them in crime series and predict where their next crimes will take place.
After identifying a series of thefts carried out in a pharmacy, for example, KeyCrime is able to statistically predict when and in which pharmacy the next robbery could be carried out. “This is the most scenic aspect, which attracts the most attention. But without the first step, without the algorithm that allows us to abstract the criminal series, we could not develop this predictive ability “, continues Venturi. Another fundamental aspect of crime linking is that it allows – after the prosecutors’ investigation work – to impute his entire crime series (and not just the event that led to his arrest) to a robber and to optimize the work of the police; making it a single office to deal with a whole series of crimes.
How many of KeyCrime‘s predictions of a future crime come true? “This is a fact that we do not provide and that I personally consider unimportant,” says Venturi. “What matters is the reduction of crimes, which has reached 50%. A percentage that represents a unique case in Italy and that is confirmed by independent analyzes based on Ossif data “.
However, the use of crime linking has another important implication: “We do not criminalize areas, as can happen with the use of hotspots. We focus only on the perpetrator or perpetrators of the crimes we have identified and the targets at risk “, explains Venturi. An important clarification, which can greatly reduce the danger that these tools are transformed into a technological go-ahead to subject to indiscriminate controls on those who may have the sole fault of living in less safe areas of cities or of walking in a neighborhood highlighted by an algorithm.
KeyCrime, unlike other systems (which have renounced it following numerous controversies, which have shown how the use of these algorithms can cause discrimination against the weakest subjects in society), also archives and uses data relating to ethnicity of the robber. “In terms of investigation, information on the ethnicity of those who committed the crimes is essential; if developers of other software have decided not to collect this data, I have serious doubts about the effectiveness of their software.
The risks of predictive policing
To clarify this fundamental point, it is useful to give an example. A software like PredPol could indicate that in a district of Milan, at a certain time and based on other factors (including weather conditions), it is possible that violence may have occurred at the hands of a certain ethnic minority. This system risks causing a redundancy of checks and searches on individuals who are solely guilty of being of the same ethnicity as that indicated by the software. An unbearable consequence (which, for obvious reasons, cannot instead occur against the ethnic majority) that led some software to exclude information relating to ethnicity from the database.
KeyCrime, working on the profile of the authors of a criminal series and on the precise places where they could strike again, effectively reduces this risk; providing much more detailed support to the police than an indication of which area to patrol. “We know many characteristics of the person we are looking for,” continues Venturi. “We know, for example, that runs on a white scooter, which is 1.82 tall, powerfully built and also what ethnicity he is. In some cases, we also have a picture of him. As a result, the police do not stop people indiscriminately, but only those who match the profile. Furthermore, robbers are usually blocked just before carrying out the robbery, when they are already with the gun in their hand, or even immediately afterwards “.
Despite valid reassurances on the functioning of Venturi’s algorithm, a doubt remains: would it not be the case that an algorithm with tasks of this responsibility, and which handles such sensitive information, were transparent and can be analysed by public opinion? “There is no need,” concludes Venturi. “Prevention in our case is targeted: the software only offers you crime linking and prediction of the next objectives. KeyCrime also has no scientific value in court. Our algorithm is part of a process that follows a normal judicial process and where the police decide whether and how to use the information we provide it. KeyCrime is an important part of an investigative process, but the last word must always be that of the investigator. ”
Source: Wired Italia