Featured Image: Etienne Girardet
Picture this. You haven’t broken the law in any way, you live in a certain area and you’re friends with certain people. This life you have and these particular circumstances you’re in are something that an algorithm finds statistically worth flagging.
Before you’ve done so much as jaywalk or double park, the algorithm has decided that you’re a person of interest.
This is the kind of reality that predictive policing brings into the picture. The future is pre-judged and the concept of presuming one innocent is a quaint relic that the software wasn’t trained on.
Predictive policing is the use of AI and data analysis to forecast criminal activity before it happens, or identifying individuals that are seen as likely to break the law. Law enforcement across different countries has been quite receptive to these tools.
Despite the very enthusiastic response from police departments, the ethical problems posed by this technology seem to be multiplying faster than the actual crime rates it was meant to reduce.
What Does “Garbage In, Garbage Out” Mean in This Context?
These algorithms are trained on historical crime data, and unsurprisingly, historical crime data is not a neutral record of reality. It just reflects who got policed.
The OHCHR’s Special Rapporteur on racism found that predictive policing tools that are based on location create a feedback loop. Officers over-police certain neighborhoods which leads to them recording new offenses there. The algorithm then doubles down and produces predictions that are increasingly skewed and target the same neighborhoods.
Past bias inevitably leads to future bias.
The NAACP has detailed how reliance on historical criminal data is compromised from the get go because black communities or poor communities have been the targets of disproportionate over-policing for decades. Using the same data to train AI doesn’t legitimize it.
There is an additional part that people do not like talking about: sometimes the data is deliberately distorted. Research published on arXiv noted cases in the U.S. (Louisiana and Connecticut) where police officers deliberately lied about the race of people that were stopped or cited for the purposes of covering up discriminatory behavior.
New York’s Stop-and-Frisk Problem, Now With Algorithms
Research published in 2025 found that the highest volume of stop-and-frisk scenarios since the year 2014 were recorded in 2024. Nine out of every ten people that were intercepted were black or Latino. Over 69% of those stops were just dead ends that didn’t reduce crime. All they did was inconvenience certain groups of people.
The Software Doesn’t Even Work Well
The bias issue might have been worth tolerating if the technology was actually extremely accurate, but it’s not. The Markup’s investigation into PredPol (now Geolitica) assessed more than 23,000 predictions that were generated for the Plainfield, New Jersey Police Department. The success rate was under 0.5% with robbery and assault predictions hitting 0.6% and burglary predictions managing a barely-there 0.1%.
The company went out of business in 2023.
The Global Issue
The same ethical failures we just highlighted are happening across Europe, Asia, and Africa. In some cases, those failures are amplified.
The Ministry of Justice in Britain is working on something that most describe as a murder prediction tool.
How does it work?
According to freedom of information disclosures obtained by civil campaigners, the tool will tap into data about individuals’ addiction records, mental health histories, and recorded incidents of self harm. I think we all agree that this type of information has no valid place in a pre-crime calculus.
India’s Maharashtra state, on the other hand, has been taking strides to expand its MARVEL system. The system shares data collected by the police with a state-owned AI firm and a private company for the purposes of autogenerating case files and enhancing intelligence.
The Carnegie Endowment for International Peace has noted that authoritarian governments in the Gulf, East Asia, and Central Asia are the number one adopters of AI surveillance tools. It’s also interesting to note that liberal democracies in Europe have also been racing to adopt their own predictive policing and facial recognition systems with way less scrutiny than the topic deserves.
Africa’s case is an example of what happens when surveillance technology arrives in the absence of an adequate legal framework to govern it.
Under the banner of “safe city” and “smart city” programmes, governments across the continent have been deploying AI-powered surveillance systems, primarily financed and built through partnerships with Chinese firms. Huawei’s Safe Cities project furnished Nairobi and Mombasa with 2,000 CCTV cameras; Uganda’s Kampala received a $126 million system with 1,800 cameras and facial recognition capability; Zimbabwe’s Zim Cyber City project involves Chinese firms including Hikvision and CloudWalk Technology.
A 2025 investigation found that these smart city projects, presented publicly as crime-fighting infrastructure, have in practice been used to surveil journalists, opposition figures, and political opponents, with spyware like Pegasus allegedly deployed during election cycles in Kenya.
In Nairobi, crime rates reportedly increased after the Safe City project was implemented, which is a detail that tends not to feature in the vendor brochures. ENACT Africa has noted that across Africa, private security companies are typically far more technologically advanced than the police forces they operate alongside, many of which lack basic internet access and data infrastructure. Into that gap come foreign vendors, usually without democratic oversight or independent evaluation.
The continent’s weak or absent data protection frameworks mean that algorithmic systems trained on corrupted, colonial-era policing patterns can be deployed at scale, with no legal mechanism to challenge their outputs. Africa imports the technology and absorbs the consequences, while the intellectual property, the profits, and the accountability remain elsewhere.
Transparency? What Transparency?
One particularly frustrating feature of the predictive policing industry is its opacity. These are proprietary systems: private contractors sell black-box tools to public agencies, and neither the affected communities nor their elected representatives have meaningful access to how decisions are made. The NAACP flagged this directly, noting that the proprietary nature of these algorithms prevents any public understanding of the decision-making framework. You can be surveilled, flagged, and policed by a formula that no one outside a corporate legal team has read.
The Electronic Frontier Foundation has pointed out that the predictive policing market has been consolidating around a small number of vendors, with SoundThinking (formerly ShotSpotter) having already absorbed both Hunchlab and Geolitica.
When harmful and flawed technologies bundle together, municipalities that buy one product often end up deploying a full suite, without meaningful elected oversight or public accountability.
The Presumption of Innocence, Quietly Retired
Maybe the sharpest ethical edge here is the one that gets the least attention: the philosophical rupture at the heart of predictive policing. Western legal tradition is built on the idea that you are innocent until proven guilty. Predictive policing inverts this. It treats the probability of future guilt as sufficient justification for present-day police attention. The ACLU has described this as creating a presumption of guilt by association, in which individuals and entire neighborhoods are flagged based on statistical inference rather than observed behavior.
Predictive policing’s person-based variant makes this even more explicit: a score is assigned to an individual, not a place, and that score shapes how police interact with them. Chicago’s Strategic Subject List reportedly assigned scores to 56% of Black men aged 20 to 29 in the city, according to a Chicago Magazine analysis. Not because of what they had done, but because of where they lived, who they knew, and what the algorithm decided those factors meant. That is not law enforcement. That is bureaucratic pre-crime.
What Would Responsible Use Look Like?
Not everyone calls for an outright ban, though the EFF and several cities have done exactly that. More moderate voices call for independent oversight of algorithms, mandatory disclosure when these tools are in use, and audit mechanisms to review whether outputs are fair and accurate. The NAACP recommends establishing rigorous, independent oversight bodies and requiring law enforcement agencies to disclose not just that they use predictive tools, but how those tools work and what data they consume.
The legal landscape is also shifting. The Johns Hopkins University Law Review noted that existing equal protection frameworks require proof of discriminatory intent, not just discriminatory impact, making algorithmic bias difficult to challenge in court under current standards. Updating those frameworks to account for statistically demonstrable harm may be the only way to impose meaningful legal accountability. The law, in other words, has a software update pending.
The Honest Conclusion
Predictive policing promises science and delivers prejudice with extra steps. The tools don’t work well, the data they train on is structurally compromised, the companies selling them have limited accountability, and the communities absorbing the consequences tend to be the same communities that have been absorbing consequences from the criminal justice system for a very long time. An algorithm that can’t predict burglaries with better than 0.1% accuracy is not a crime prevention tool. It is an expensive way to automate a pre-existing bias.
If AI is going to play a role in public safety, it will need to be built on clean data, governed by transparent oversight, and subject to meaningful civil rights scrutiny. Until then, the most accurate prediction these systems produce is that the people watching will be the same ones who have always been watched.
Leave a Reply