Facial recognition in Zurich
Security gain or the beginning of mass surveillance?
I grew up in a system where, even as a child, you learned what you were allowed to say and what was better not to say. My parents often said to me: “Be careful, you must not tell anyone about that.” At that time, the Ministry for State Security, the Stasi, operated the largest surveillance apparatus in world history. In 1989, around 280,000 people were officially employed in police-related roles, plus another roughly three million citizens in security-related functions. If all these forces were added up, the “security density” amounted to one operative for roughly every five inhabitants. Control was not a feeling, but everyday life. You paid attention to who was listening. You did not wonder whether you were being watched. You knew it. That was many years ago. Today I live in a democracy, in a system of rights, freedom of expression, and transparency.
In the last days of November 2025, the Zurich Cantonal Council took a decisive step: it decided that automatic facial recognition in public spaces should be permitted in the future under certain conditions. This decision immediately triggered an intense debate. Supporters speak of an opportunity for more security, for example in the search for wanted individuals, in cases of violence in the city center, or when dangers threaten at large events. Critics, however, warn of a dangerous breach of the dam: left-wing parties and data protection advocates speak of the danger of “comprehensive surveillance.”
I know what it feels like when the state comes too close. When you feel small. When you are not sure whether you are allowed to speak. I despise it to this day, and that is precisely why it is important to me that it never happens again.
But just as dangerous as mistrust is helplessness. When crimes remain unsolved. When violence in public spaces increases. When the police can no longer keep up. Then people lose trust in the state. In the context of this article, the technology of facial recognition is explained using the example of the Vaidio AI Vision platform. It is not only about the technical functionality, but also about application and benefit.
What is the initial situation in Zurich
In Zurich, as in many large cities, the police face a growing number of crimes and a lack of resources on a daily basis. The 2024 police crime statistics show: Zurich is increasingly under pressure. In the city area alone, more than 48,000 offences were recorded, around 4.5 percent more than in the previous year. In the entire canton of Zurich, more than 110,000 offences under the Criminal Code (StGB) were recorded.
By far the largest proportion of all offences concerns property crimes, accounting for over 66 percent of all cases. These include theft, burglary, fraud, and property damage. But especially here, the clearance rate is low: two out of three cases remain unsolved.
An increase can also be observed in the area of violence. In 2024, more than 10,700 violent offences were recorded. What is particularly striking is that most of these offences do not occur in private spaces, but in places where people come together. Public space forms the main focus: streets, squares, parking lots, gastronomy, retail areas, educational institutions, or leisure locations are the scenes of spontaneous conflicts and escalating situations. More than two thirds of all acts of violence in Zurich occur in these areas, often within seconds and usually without warning. The arrival of the police at the scene almost always occurs afterwards, because preventive intervention is very difficult. A structural problem further exacerbates the situation: compared to European countries, police density in Switzerland is significantly below the recommended ratio of 1 to 300 inhabitants. Current estimates for 2025 indicate 1 police officer per 477 inhabitants. In a European comparison, Switzerland is one of the countries with a very low police density.
A further look at the statistics also reveals a decisive finding: crime is not evenly distributed across the population. In 2024, 16,751 adult persons were recorded for criminal offences. Of these, almost half committed only a single offence. The other half, however, attracted attention multiple times. More than 50 percent of the investigative work was therefore accounted for by persons who committed two, three, or even more offences within a single year. A small group of repeat offenders is disproportionately active and accounts for a significant part of the overall situation. This makes it clear: not all people are potentially suspicious, but rather a relatively small, known group that repeatedly comes to attention and heavily occupies police resources.
From this insight emerges an important perspective for any discussion about new technologies. Facial recognition would not mean that all people would have to be monitored or controlled. Rather, it would be an instrument to identify known repeat offenders more quickly before they strike again.
Common myths about facial recognition
When talking about AI and facial recognition, a certain image often immediately comes to mind: cameras that see everything, know everything, and instantly identify every person on the street. A technology that observes every step, evaluates every movement, and perhaps even knows more about us than we do ourselves. This idea alone is often enough to trigger strong emotions. Fascination on the one hand, concern and distrust on the other.
Media reports also shape the image: scandals and possible abuses move into the spotlight and reinforce scepticism. Prominent reversals and debates at the EU level (campaigns against mass biometrics) often inspire doubt rather than trust. Overall, the public perception often conveys an image of an “all-seeing AI” without privacy.
In our everyday lives, however, we use facial recognition completely voluntarily, without feeling pursued. A typical example is unlocking our smartphone. When we look at our phone in the morning and unlock it with a glance, we do not perceive it as surveillance but as protection and convenience. We use the same technology only in a safe, private setting that we control ourselves. There, facial recognition does not appear threatening but helpful.
That is precisely why it is necessary, before talking about opportunities or risks, to first understand the technology. How does an AI “see” a face? What does it actually recognize – and what not?
What is facial recognition and how does it work?
Face recognition like that of Vaidio is based on deep learning algorithms. In this process, faces are first detected in the video image, aligned, and converted into numerical feature vectors (“embeddings”). This process is typical for modern FR systems: from the image data, automatic feature extractions are generated by a neural network, providing robust, semantic descriptions of the faces. The system then compares the captured faces with a database of known templates. If a match is detected (threshold-based), an identity or a group or role assignment is returned.
To better understand how facial recognition works, it helps to look at two different situations. Let us imagine a place, for example a train station or the entrance to a building. A camera records people passing through. The software behind it automatically recognizes that a face is visible and begins to analyse it. Now everything depends on whether the person has been stored in a database or not.
If the person is not stored in a database, very little actually happens. The system notices that a face is in the image, creates a short-term “faceprint,” meaning a mathematical pattern based on certain features such as eye distance or chin shape, and then discards it again. For the AI, this person remains anonymous. He is not recognised, not identified, and not stored. The AI therefore sees that a person is there, but it does not know them. One could say: the face is only a neutral image, without a name and without meaning.
It is different with a person who has been intentionally stored in a database, for example because they are on a wanted list. If this person appears in front of the camera, the software again analyses their face and creates a faceprint. This is compared with the stored codes in the database. If a match is found, the system recognises the person and can trigger an alert immediately. One could say: the AI remembers them, not because it knows every person, but only because they were deliberately stored beforehand.
How investigations are conducted today
First, video recordings from surrounding cameras must be secured and then reviewed. These are often hours or even days of video material, just to find out when the perpetrator appears, where he went, and whether he had been at the location beforehand. If there are several cameras, movement profiles must be manually pieced together. This means: investigators spend days trying to find out whether the person reappears at certain locations. Often, suspects must be observed, movements documented, and behavioural patterns collected. Such measures tie up a lot of personnel and usually take place in the background, sometimes successfully, sometimes without any result.
How AI-based video surveillance with facial recognition changes investigations
With an AI-based video analytics platform like Vaidio, this entire investigation process changes fundamentally. As soon as video material is available, regardless of whether it comes from public cameras, private systems, or a shop, it can be analysed automatically and searched for specific characteristics. The system detects, for example, a person with a hood, estimates gender and approximate age, records clothing characteristics, noticeable accessories such as bags or backpacks, and analyses movement patterns. This information is not limited to a single camera. They are, if authorised, linked across all connected cameras. Instead of manually going through countless hours of recordings, the system provides answers to specific questions: When did this person appear for the first time? Was this person already present at the same location on previous days? Is this person travelling alone, or do they repeatedly appear in the vicinity of the same group of people?
If facial recognition is used in addition, the process changes even further. In this case, the previously generated faceprint of a person is compared with stored entries in a database. If a match is detected, this does not simply mean a hint but a clear location reference: the system shows when and where a wanted person was last seen, for example at a train station, in the old town, or near a crime scene. This turns an anonymous video sequence into a concrete lead. The police not only know that the person was on the move, but also where they were, in which direction they went, and whether they had previously appeared at another location. And that in a matter of seconds. The special thing about it: not only does the search time shorten drastically. The investigative approach also changes. Observations no longer need to be reconstructed laboriously, but can be traced. Suspicion is no longer discovered by chance, but made visible. Instead of starting investigations from zero, they can begin where previously investigators arrived only after days.
A brief look in the wrong direction
A common objection is: “We already have all biometric data in passports – so such a database already exists.”
Yes, it is true: in Switzerland, biometric data for passports is stored centrally in the Identification Documents Information System (ISA). It contains personal details, the passport photo and fingerprints, exclusively for administrative purposes. The purpose is clearly limited: identification in official procedures, support in case of loss reports, protection against identity fraud. Only authorised authorities such as Fedpol, cantonal authorities, or the border guard corps have access. But what is crucial is what ISA is not: ISA is not a surveillance system, not a live database for cameras, not an instrument that automatically identifies faces in public spaces. But it would only be half a look in the wrong direction if one were to stop there. In theory, a national facial database could be created; technically, this would be feasible. But for this, millions of biometric data records would have to be collected, validated, and be retrievable in real time. If every face in these recordings were compared with millions of biometric profiles, an enormous volume of data would be generated that could only be handled with a national high-performance infrastructure.
Such systems do indeed exist. In countries like China, data centres are part of the state infrastructure and linked to nationwide camera systems. When we look at China, we see where such a development can lead. There, large urban surveillance systems are directly connected to centrally managed biometric databases. Cameras are designed for real-time recognition, and movements in public spaces can be automatically assigned to personal profiles. Facial recognition there is part of a comprehensive state infrastructure that is used not only for security but also for administration, access control, and in some cases even for social rating systems. This model is based on different legal and societal foundations, on an understanding of the state and the citizen that is not compatible with European principles of data protection and freedom, and hopefully never will be in the future.
Ensuring that this never happens again is our responsibility. Not in the responsibility of any software, but in the strength of our constitutional principles. In our laws, in our transparency, in the obligation to provide justification, and in the right to object. Democracy is not something that you achieve once and then keep. It is defended every day, especially where new technologies emerge. If we define clear boundaries, build in protective mechanisms, and keep control with humans, facial recognition can be part of a modern security architecture. Not as a tool of power, but as a service to society.
Conclusion – Relief, not control
With modern facial recognition, it is not about capturing all people, but about quickly recognising a few known individuals who repeatedly appear. Crime data show that the general population would not need to be the focus, but a small portion of repeat offenders. This is exactly where intelligent video analysis can help.
That is a crucial difference. The term “mass surveillance” often appears in debates, but technically and legally, comprehensive facial recognition would be neither sensible, nor practical, nor compatible with democratic values. It would waste resources, generate volumes of data that would be difficult to process, and ultimately lose sight of what is truly relevant.
Instead, a system like Vaidio pursues a clear goal: to prepare information in such a way that investigators can start where it matters. It represents a gain of time for the moments in which every minute counts.
Modern security does not mean: seeing everything.
Modern security means: recognising the right thing.
