WritingResources
Back to all posts

Urban Digital Twins and the dawn of predictive policing

Ludivine Siau

The problem

Planning law enforcement operations, like managing traffic and crowds during protests or responding to a crisis, is a nightmare of moving parts. Authorities rely on table-top exercises using historical crime analysis and static maps, which fail to account for real-time dynamics of a city, while physical drills are resource-intensive and logistically difficult, and sometimes limited by safety concerns.

The proposal

The paper proposes an Urban Digital Twin (UDT): a high-fidelity virtual mirror of the city that uses real-time data and dynamically responds to simulated actions:

  • Urban GIS data, real-time sensor feeds, mobility patterns and crime logs are fused into a unified “urban state encoding”
  • The framework uses Agent-Based Modelling to represent civilians, offenders, and police units and predict their “plausible actions” in response to the environment and other agents.
  • AI allows commanders to run hundreds of “what-if” simulations and uncertainty analyses to balance multiple goals, such as minimising response times, maximising safety and fairness, while adhering to legal and ethical constraints
  • A continuous feedback loop, calibrated on historical data, keeps the state of the UDT up to date.

The results

The researchers compared the proposed digital twin-based planning against traditional methods:

  • The average police response time was slashed from 12.8 minutes (baseline) to just 6.1 minutes
  • Effective monitoring of high-crime areas increased from 62% to 84%
  • The rate of successfully de-escalating or preventing incidents rose from 58% to 81%
  • Patrol allocation efficiency improved by 13%

The paper concludes that the framework would be especially valuable for training purposes, to simulate rare and high-stake situations that cannot easily be mocked in a physical exercise.

Caveats and can of worms

The core limitation here is the validation setup. The study measures success within its own simulation. Until these results move from a "digital sandbox" to a live city pilot, the improvements are theoretical and might be biased in favour of the AI. We would need to see how this system integrates with end-to-end human planning and decision-making workflows in law enforcement. Does it act as a pre-event planner (e.g. simulating 10,000 London marathons) or a tactical real-time navigator that predicts suspect movement based on live data? How would it influence human behaviours, social processes and even criminal intent?

This technology promises many benefits for law enforcement: informed decision-making, resource optimisation, cost reduction, safety improvements. But it brings several systemic risks that the authors do mention but leave for future studies:

  • Reliable accuracy: modelling human behaviours and processes at the level of big, crowded cities is still a frontier capability. If virtual models do not fully reflect reality, actions taken based on their predictions could lead to unintended outcomes.
  • Legal and ethical boundaries that protect privacy, safety and civil liberties: to function reliably, the UDT requires massive data ingestion of CCTV, mobility traces and other public infrastructure feeds. Where do we draw the line between public safety and total surveillance?
  • Monitoring, governance and explainability: the authors claim this is a "decision support" tool that shouldn’t be used to automate law enforcement. But when the system tells a commander they have an 81% chance of success if they deploy a riot squad there, the suggestion carries a lot of weight. We must account for automation bias. Will a human feel comfortable overriding the AI-optimised solution? How will limitations be communicated to its users? For this to be ethical, the system must explain why it made a recommendation in a way that allows a human to spot a logic flaw before actions are taken and processes must be in place to audit the use of the system and downstream decisions.
  • Fairness: if the UDT is calibrated on historical crime data, which we know is often biased, the AI might simply predict more crime in the same neighborhoods. This leads to more patrols and more arrests, creating "new" data that justifies the original bias. It’s an automated self-fulfilling prophecy. Same problem as most AI systems, except this one is especially intricate. I wonder what measures of fairness were applied during the study’s experiments.
  • Cybersecurity: Hosting a city’s entire safety strategy in one Digital Twin creates a single point of failure. The ultimate honey-pot for cyberattacks. Data manipulation, infrastructure vulnerabilities, DoS attacks… I wouldn’t want to be in charge of securing a system that, by definition, must be connected to every sensor in the city and on which authorities rely in situations of crisis.

Real-word experiments

AI-supported law enforcement has already hit the pavement, with a mixed track record at best. To understand the risk of a full Urban Digital Twin, we have to look at the simpler systems that preceded it.

We cannot talk about “predictive policing” without mentioning the landmark 2016 ProPublica investigation into the COMPAS software. That study found that the algorithm used to predict recidivism was twice as likely to falsely flag Black defendants as future criminals compared to white defendants.

The Netherlands had its own brush with biased algorithms in 2020 when Amnesty International sounded the alarm on the Sensing Project in Roermond. Designed to stop "mobile banditry" (pickpocketing and shoplifting), the system used remote sensors to track vehicle movement patterns and produced risk profiles that allowed police to make targeted interceptions. Amnesty International warned that the AI’s predictions appeared to be biased towards certain Eastern European nationalities.

If we couldn’t get neutral policing right with systems that are far from the complexity of a full UD, can you imagine how impossibly complex it might get with more data sources, interactions and feedback loops?

The Dutch are now moving toward a high-fidelity Urban Digital Twin: PULSE (Proactive Urban Livability and Safety Engine). Funded by the European Urban Initiative, the project launched last year in Heerlen, a city struggling with high crime and social deprivation. The goal is to map the city’s “social heartbeat” by fusing security data with community input to replace guesswork with evidence-based policy.

PULSE is currently in its alpha-beta phase, with field testing for an AI co-pilot set for December 2026. The developers admit that integrating diverse datasets while respecting GDPR is a “delicate balance”.

Let’s wish them the best and assume, for a moment, that the technical execution is flawless. I still have a fundamental objection.

I understand the instinct to hoard data, the hope that if we just gather enough variables, a pattern will emerge that finally allows us to make the "right" decisions. But there’s a real risk of data dredging with these systems: if you torture a massive dataset long enough, it will eventually confess to a correlation. Too often, it’s just mistaking statistical noise for significance.

Evidence-based policy is a great goal, but there is a massive leap between measuring the impact of past initiatives and simulating the impact of future ones. Big Data and AI have become technocratic shortcuts to avoid the difficult work of designing transparent policy experiments and committing to rigorous evaluation. They also serve as excuses to stop engaging with residents as people and start managing data points instead.

My fear is that predictive tools applied to policing and policy will move us toward a world where the system feeds itself and where changing the status quo becomes impossible. Governance is not an optimisation problem to be solved by a machine. We still need policy-makers who are critical thinkers and who are willing to bring new ideas to the table instead of delegating the moral weight of tough decisions to an algorithm.

References

V. Goel, R. Ranjan, M. Kosimova, S. Murodjon, X. A. Toirovich and A. Reymbayev, "Digital Twins for Simulation of Law Enforcement Operations in Urban Environments," 2026 5th International Conference on Innovative Practices in Technology and Management (ICIPTM), Noida, India, 2026, pp. 1-6, doi: 10.1109/ICIPTM69057.2026.11465731

"PULSE-TWIN - Proactive Urban Livability and Safety Engine with Integrated Social Digital Twin", European Urban Initiative (date visited: 07/05/2026)

M. Koning, "Netherlands: End dangerous mass surveillance policing experiments", Sep 29, 2020, Amnesty International (date visited: 07/05/2026)

J. Larson, S. Mattu, L. Kirchner, J. Angwin, "How We Analyzed the COMPAS Recidivism Algorithm", May 23, 2016, ProPublica (date visited: 07/05/2026)