While Dick Willems is coding on a computer in a small office at the headquarters of the Dutch police in Amsterdam, a sense of chaos and excitement lingers in the air. It is 2013 and the Dutch police is going through a major transformation: 25 regional units are merged into a single National Police. Key to the process is a new approach to ‘intelligence-led policing’, a strategy which aims to use data and knowledge to deploy police resources more effectively and plan incident-driven approaches proactively instead of just reacting to crime.

Dick Willems is a seasoned data scientist. He was hired to develop a ‘predictive’ policing algorithm for so-called “high-impact crimes” such as burglary. The system is inspired by an existing model which he developed earlier in his corporate career: Originally, this algorithm was meant to predict which customers were most likely to switch away from certain brands.

In 2015, Willem’s work at the police resulted in the birth of CAS, short for Crime Anticipation System. The goal of CAS, much like similar statistical tools tested in different cities and regions across Europe, was to classify crime-prone areas across the country and allocate police resources accordingly. CAS was initially created to operate in Amsterdam only, but the system would  become the norm for ‘predictive’ policing in the entire Netherlands for years to come.

The Dutch police regarded themselves as the pioneers of ‘predictive’ policing in Europe, since they used CAS in the whole country and not just a few cities. Until late 2025, when the once-so-beloved system was abruptly decomissioned. The reason: its “added value” was “unclear”, as a devastating internal police report that was carried out a decade after the system’s deployment exposed. The report has now seen light of day due to freedom of information requests.

As stated in this ‘Kraai’ report (Quality and Risk Management System for Algorithms and AI) from September 2025, only one out of 50 incidents in the city of Amsterdam was correctly predicted by the algorithm. This represents 11,5% less than what police researchers claimed in the early stages of the system’s deployment eight years prior.

Confronted with this number, the police admits that they stopped using CAS “because it was impossible to prove that burglaries decreased due to the use of the system”, a police spokesperson admitted via email. The story of how they ended in this tight spot is even more convoluted.

Hot times and hotspots

The Crime Anticipation System divided cities into a grid of areas each measuring 125 square meters. Subsequently, it collected as much information as possible about each area. For instance, how many crimes occurred in the past or whether known suspects lived nearby. The system combined police information with data from the Central Bureau for Statistics. This allowed the police to add socioeconomic data to the formula, such as the average income of a neighborhood or whether large families lived there.

The system assessed burglaries, car or bicycle thefts and nuisance. The output were the so-called ‘hot times’ and ‘hotspots’.  Reading the system’s recommendations was pretty straightforward: specific areas on a map of the city were automatically colored if CAS recommended deploying police patrols during time intervals consisting of four hours each. CAS is a closed system, which meant that police officers could not see what data the model was using to make a specific prediction.

“Practically speaking, the algorithm is actually very simple,” says researcher Lauren Waardenburg during a conversation via Zoom. Waardenburg followed the CAS data team and police officers for several years as part of her work at the Free University of Amsterdam and ESSEC Business School. “Officers are assigned to a small area. They then drive through that neighborhood once, maybe a second time an hour later. And that’s it. If they are there for five minutes, that’s already a lot.”

Living close to a burglar

CAS considered certain areas more dangerous than others based on the data it was trained on, including the number of criminal incidents taking place in a location, or even the physical distance to the closest known offender (called ‘suspect’ from 2017 onward).

“The reason these predictive policing programs often focus on burglary is because this crime type has almost a 100% reporting rate”, explains Christoph Vandeviver, Research Professor at Ghent University: “People have to report the crime to the police to receive compensation from their insurance.” Also, the location of a home is fixed, he continues. “This means there’s little room for mistakes in our spatial data points”.

However, the specific timing of a burglary is less clear. People are often either not at home or asleep when the burglars come in, leaving a varying time span during which the actual crime could have occurred. In CAS, this uncertainty was dealt with by using the exact medium between the time someone left their home or went to bed and the time the burglary was first noticed. According to a paper by Dutch researchers Serena Oosterlo and Gerwin Van Schie, this represents a limitation to the predictive model.

Besides crime data, sociodemographic variables were also vital to CAS’ predictions. The system looked at the size of families and their income, the number of social benefit recipients, as well as the number of single-person households in an area. Researchers Oosterlo and Van Schie write that CAS frames potential burglars and/or victims as members of “broken families” or with poorer backgrounds.  The indicator on the number of non-Western immigrants, present in the initial list, was removed in 2017 because, according to Willems, it did not add any predictive value.

This investigation revealed that in 2023, the same year the police conducted a Human Rights and AI Impact Assessment, all socioeconomic data points were removed due to their “limited effects on the algorithm’s predictions”. The analysis showed that the use of CAS could impact at least four individual rights: the autonomy and presumption of innocence, the prohibition of unequal treatment as well as the right to explainability of the algorithm.

Implementing the system nationwide

Two years after the first experiments, the police decided to broaden the scope of the system beyond Amsterdam. In 2015, CAS pilot projects were carried out in the police departments of Enschede, Groningen-Noord, Hoefkade and Hoorn – smaller localities distributed across the Netherlands. These projects were then evaluated by three researchers affiliated with the Dutch Police Academy, which combines a police school and a research department. 

Back then, the Academy researchers assessing the pilot projects admitted in their report that they also felt “the enthusiasm for predictive policing”. They calculated that CAS correctly predicted about 13,5 % of the areas in which burglaries took place. That is, five times better than the areas identified by randomized patrols. 

However, they also found that there is no evidence that makes the “supposed effects” of ‘predictive’ policing “plausible”. More specifically, the researchers did not find a significantly lower crime rate in the four police departments that had used CAS for their patrol deployments, compared to the rest of the country. This means that CAS had not delivered on its promise of dwindling crime. 

On the other hand, police officers were mostly not convinced by the promise of ‘predictive’ policing. Among CAS users, a greater share believed that the system failed to improve effectiveness, team capacity allocation, or job satisfaction than those who believed it did. 

Regardless of the pilot evaluation, the police announced that they were planning to implement CAS nationwide in May 2017. Back then, the Dutch police bragged about being the only European country using a ‘predictive’ policing software at the country level.  

“After the report by the Police Academy was released, all the warning signals were there”,  recalls Stijn Ruiter, Professor of Evidence-Based Policing at Vrije Universiteit Amsterdam.  “There was so little scientific evaluation for this program, and still the police wanted to proceed with it.” 

Analyzing car thefts in zones without cars

The national roll-out began in 2017, starting in more than half of the then 167 Dutch police departments. Researcher Lauren Waardenburg followed the data team in Amsterdam and the police officers in the field, who often seemed bored of using CAS. “Patrolling these grids is quite absurd for the officers”, she recalls. “The whole idea is that there is no spike in crime precisely because you are present in a location. That can feel very aimless for the police.” 

Especially when the model started blurting nonsense. For instance, CAS suggested officers to patrol for car theft in car-free zones. At other times, the model stopped producing predictions altogether because according to the police itself and Waardenburg’s research, there was not enough crime data to feed it. Criminologist Stijn Ruiter says these problems reveal a major shortcoming of all ‘predictive’ policing models: “What are police officers supposed to do with those predictions in the field? The focus is always on the accuracy of the prediction, but the ‘policing’ itself has received little attention.” 

By 2022, the Dutch Court of Audits also expressed severe apprehensions. The independent body screened nine government algorithms used in different domains. They applied 12 criteria divided in four larger categories: accountability, bias in the training data, privacy and IT management.

CAS performed worst of all systems tested: it scored mid to high risk on every single criterion. The Court of Audits wrote that there was insufficient control over biased outputs, that no privacy analysis was performed, and that the police collected more data than necessary. 

Yet again, the police did not seem very impressed by the Court’s analysis. In a letter, the police wrote that suggesting that CAS posed a high risk was “unjustified”, and that a privacy assessment was therefore “not necessary”. The algorithm could thus “continue to be used”.

The (unexpected) end

In June 2025, just weeks before the Kraai meeting would be held, the police’s Scientific Advisory Council published yet another critical report. Albeit more general in scope, it stated that the police should use their intelligence for retrospective and real-time work rather than trying to predict the future.

In line with this overview, the Kraai report concluded that CAS should be discontinued. Regarding the poor ‘predictive’ capacity, the internal report also addressed that the ultimate goal of the system was actually not defined properly. Along these lines, they claimed there were “no formal norms” regarding when teams should use CAS, creating a risk that money is spent maintaining a system “that is barely used.”

Police officers using CAS were also not informed about the limits and risks of the system, the report stated. Police and intelligence officers may therefore overestimate the model, leading to so-called automation bias. On top of all, the experts write that CAS is “possibly used on unjust judicial grounds”, meaning that the data processing might be unlawful. 

Finally, in January 2026, the Algorithm Board of the Dutch Policy decided to put a halt to the system, alleging alternatives to CAS “seem to be more effective and contain fewer risks related to the data used”.

“Public conscience about the risks of these AI-systems has increased a lot in the Netherlands after we had some big scandals here”, says Marc Schuilenburg, a professor of Digital Surveillance at Rotterdam University and co-author of the Scientific Advisory Council report. The Dutch government fell in 2021, after SyRI, an automated benefit system, falsely accused thousands of parents of fraud and devastated their finances. More backlash arose against the use of a controversial algorithm that aimed to predict which individuals may showcase violent behavior. These cases created public outrage and a debate about the need for oversight of ‘black box systems’. 

Big data policing plans

With the quiet withdrawal of CAS, however, predictive policing in Europe is still far from its final days. The British government announced that by 2030, it plans to develop a “detailed real time and interactive map” that can predict where knife crime will occur. It also aims to record “antisocial behavior, before it spirals out of control”, which implies predictions at an individual level.

In parallel, a team of criminologists at the Ghent University in Belgium have developed an AI model to predict residential burglaries. This ambitious project called Bigdatpol is intended to become the benchmark for big data policing in Europe.

Over the past years, the Bigdatpol team has already been in contact with police departments and companies in Austria, Germany, Hungary, Serbia, Spain, the UK and, almost ironically, the Netherlands.

CAS creator Dick Willems declined our invitation for an interview.

Our Freedom of Information Request to the Dutch Police did not give us access to the most recent variables list used in CAS. 

This story was produced in the framework of the Algorithmic Accountability Reporting Fellowship by AlgorithmWatch, and the support of the Pascal Decroos Fund for Special Journalistic Projects.