Torrejón tests AI to read safety patterns, not to watch people

On June 3, 2026, the city council of Torrejón de Ardoz presented Polic‑IA at South Summit Madrid 2026, a tool that maps where and when historical data suggests greater risk. It is not the same thing as a surveillance camera, a suspect list, or a system that makes decisions for an officer.

According to the municipality’s announcement, the web tool combines machine learning and geospatial analysis to study incidents recorded by the Local Police. A user can choose a location, date and time slot; the application returns a risk estimate and makes it possible to review the historical concentration of incidents and compare scenarios. The initiative was developed as a final-degree project by Álvaro Juárez García, a Computer Engineering student at UCAM.

Forecasting patterns, not people

The council explicitly sets limits around the project: Polic‑IA does not identify people, carry out automated surveillance, make autonomous decisions, or claim to determine where a specific crime will be committed. Its stated role is to turn historical information into maps, charts and an additional signal that can help safety managers understand spatial and temporal patterns.

That distinction is essential to understanding what a tool of this kind can contribute. A map of past frequencies may help frame operational questions —such as where the allocation of resources merits examination during a given time period— but it does not establish that an incident will occur. Historical data describes what was recorded, under particular reporting practices and in a particular context; it is neither a complete picture of a neighbourhood’s safety nor a judgement about the people who live there.

The council places the estimate under professional judgement and local knowledge. That human role is not a communications detail: it affects whether the output remains preparatory information or begins to materially shape a decision. It also explains why data visualisation, traceability and performance checks matter as much as the model itself.

What the European framework asks for

The EU AI Act classifies certain uses in sensitive areas, including law enforcement, as high-risk, but it requires an assessment based on a system’s actual intended purpose. Article 6 provides a route for some Annex III uses not to be treated as high-risk when they do not pose a significant risk and, among other conditions, only perform a preparatory task without replacing or improperly influencing human assessment without appropriate review. That assessment must be documented.

In May, the European Commission published draft guidelines to help providers, deployers and authorities apply that classification. Their practical lesson is cautious: calling a tool “decision support” is not enough. Its stated purpose, the way it is used and the documentation that makes those claims testable all matter.

From the public information available, it is not possible to certify Polic‑IA’s legal category or evaluate its accuracy from outside: the announcement does not publish the dataset, error metrics, thresholds, area-by-area testing or a review protocol. There is no need to manufacture those details to assess the announcement. Its immediate value is to illustrate a bounded use of AI: summarising historical patterns to support a planning discussion, not automating policing.

The test is everyday use

The meaningful next step is to preserve those limits when the tool is used. Publishing what data are aggregated, how they are updated, what uncertainty the model carries and how unreliable outputs are corrected would make the promise of explainable, responsible AI easier to verify. Technology can organise information that is difficult to read; decisions about resources, and accountability for them, should remain human.