Facial recognition in Zurich.Benefits and limitations.

Facial recognition in Zurich: What the debate is about

Facial recognition in Zurich is neither just a technical nor just a political question. At the same time, it affects security in public spaces, the efficiency of investigations, the protection of biometric data and the freedom to move around without being observed. It is therefore not enough to classify the technology as a security gain or as mass surveillance.

The decisive factor is the specific use: Is an individual crime scene subsequently compared with a legitimate reference database? Is a person voluntarily verified at the entrance? Or are all faces in public spaces constantly checked against watch lists? These scenarios differ fundamentally technically, legally and socially.

Four terms that should not be confused

In discussions, facial recognition, facial capture and video analytics are often equated. This creates a false picture of actual data processing. A reliable assessment begins with four clear categories.

• Face detection: A system recognises that there is a face in the image. It doesn’t yet know the person.
• Verification (1:1): A face is compared to exactly one expected identity, for example in an automated passport control.
• Identification (1:n): A facial image is compared to multiple stored reference images to suggest possible matches.
• Attribute or event analysis: Software searches for clothing, direction of movement, objects or situations without identifying a person by name.

The last point is particularly important: AI video analytics does not automatically require facial recognition. Many security tasks can be solved based on events, for example through violence, weapon, fall or intrusion detection. Where identification is not required for the purpose, it should not be used.

How automatic facial recognition works technically

The system first locates a face and uses it to create a mathematical feature vector, a so-called template or embedding. This vector is not a name. Identification only occurs when the template is compared with reference data and assigned to a data record.

The result is not an infallible determination, but rather a statistical similarity assessment. A defined threshold determines when the system reports a possible hit. A lower threshold finds more candidates but increases the risk of false hits. A high threshold reduces false positives but may miss actual matches.

Image quality, viewing angle, lighting, reference image age, occlusion, and characteristics of the trained model affect performance. Therefore, an algorithmic hit must not trigger an automatic sanction or final identification. It must be checked by qualified people based on further information.

What the Zurich Cantonal Council decided in 2026

The Zurich Cantonal Council approved the total revision of the Information and Data Protection Act (IDG) in the final vote on March 23, 2026. The decree does not categorically exclude pilot tests with biometric recognition systems in public spaces. However, this is not a blanket permission for widespread facial recognition.

According to the data protection officer of the Canton of Zurich, such pilot tests can only take place within the framework of the general legal requirements. They serve to test processing of special personal data for a limited period of time that still requires clarification. This requires a legally anchored task, a regulation as well as clearly defined protective measures and a careful evaluation.

In February 2026, the cantonal council also rejected an initiative by the authorities for a far-reaching ban on biometric facial recognition. This decision does not mean that every conceivable use is permitted. Purpose limitation, proportionality, legal basis, transparency, data security and control of particularly risky data processing remain decisive.

Pilot testing is not the same as continuous monitoring

A pilot should answer a clearly stated question: Is the system suitable, necessary and proportionate for a particular legal mandate? This includes a limited location, a limited time period, defined reference data and measurable criteria for benefits and risks.

An open test based on the principle “We collect first and decide later” would not be sufficient. Even before the start, responsibility, access rights, deletion periods, error tolerances, human review and information for those affected must be regulated. It must also be clear when the experiment will be ended or aborted.

The difference to Fedpol facial image matching

Fedpol introduced automated facial image matching in 2025 to complement existing biometric procedures. Facial image traces from investigations are compared with already stored images of known people. Fedpol expressly excludes automatic real-time surveillance of people through cameras in public spaces.

The difference is essential: a subsequent comparison of a concrete evidence image has a limited investigative purpose. A live system, on the other hand, would continuously record a large number of uninvolved people. Both methods use biometric characteristics, but do not create the same risk of interference.

What benefits facial recognition can have

When used in a legally limited manner, automated image comparison can help investigators sift through large amounts of images more quickly. This can be relevant if a crime scene image is the most important clue or if a wanted person is to be found based on existing recordings.

Technology can prioritize candidates and shorten search times. However, it does not replace the criminal assessment or other evidence. A responsible process treats the hit as an indication, documents the review and requires independent human confirmation.

Which risks need to be taken particularly seriously

Faces are biometric features. Unlike a password, they cannot be easily exchanged. If templates are used inappropriately, stored for too long or linked to new databases, the original purpose can gradually expand.

• False hits: Uninvolved people can mistakenly appear as candidates.
• Distortions: Performance and error rates can vary between groups of people.
• Purpose expansion: A limited database can later be used for other purposes.
• Deterrence: Perceived permanent identification can influence behaviour and the exercise of fundamental rights.
• Data outflow: Biometric templates and reference images are targets particularly worthy of protection.

Therefore, not only the average recognition rate must be considered. Those responsible need separate values ​​for false positives and false negatives, realistic testing under local conditions and documented handling of uncertainty.

Eight protection mechanisms for responsible use

1. Narrow purpose: The specific security or investigative purpose is determined in advance.
2. Clear legal basis: Responsibility and depth of intervention have been legally tested.
3. Data minimization: Reference lists only contain necessary and legitimate records.
4. Human control: No hit leads to a measure without a qualified check.
5. Short deadlines: Misses, templates, protocols and reference images are deleted in a controlled manner.
6. Technical separation: Databases, roles and systems are differentiated from one another.
7. Detectability: Queries, hit ratings, accesses and changes are logged.
8. Independent evaluation: Benefits, errors, group differences and encroachments on fundamental rights are examined.

When video analytics without identification is sufficient

Many operational situations do not require identity. If a control centre needs to know that a person is on the ground, a crowd is getting too dense or a visible weapon appears, an event-based analysis is sufficient. The system reports the situation without matching a person to a name database.

This separation follows the principle of data minimization: as little personal processing as possible, as much as necessary for the clear purpose. Facial recognition should therefore never be the default setting of a video analytics platform, but rather a separately justified function.

Conclusion: Control comes from boundaries

The question is not whether facial recognition is generally good or bad. What is crucial is which data is processed for what purpose, with what reference basis and under what control. A targeted, subsequent comparison differs fundamentally from an unprovoked identification of all passers-by.

For Zurich this means: pilot opportunities must be seen as a test assignment, not as a free pass. Democratically viable use requires strict limits, transparency, human responsibility and the verifiable possibility of switching off an unsuitable system.

For operations in the EU, the GDPR and the relevant national regulations must be checked. In the case of police authorities, special data protection law can be decisive for criminal prosecution. Swiss requirements should also be viewed as a separate legal framework. Event-based processing is not automatically anonymous. Our article provides a current overview Data protection and EU AI Act.