Biometric Categorisation

Biometric characteristics are measurable properties of the body: the face, the voice, the gait, the iris, the fingerprint. A biometric categorisation is software that captures such characteristics and then sorts the person into a group accordingly. This can be an age group, a gender, a hair colour – or something far more sensitive, such as presumed origin, religion, or sexual orientation. It is important to distinguish this from facial recognition: that aims to find out which specific person is standing in front of the camera. Categorisation does not want to know that at all. It sorts anonymous people into pigeonholes, and that is exactly where its problem lies.

When a pigeonhole becomes discrimination

Some of these classifications sound harmless. A vending machine that checks whether a buyer looks over 18 years old is an everyday case. Other classifications are dangerous because they concern unchangeable characteristics. When a camera sorts passers-by by presumed ethnicity, this quickly becomes a basis for unequal treatment.

On top of that, many of these classifications are not even scientifically sound. A religion or a political stance leaves no trace on a face. A system that claims to detect such things nevertheless produces results – just random ones that look objective. Experts often call this modern physiognomy, meaning the old, long-since-refuted idea that one could read character from appearance.

The EU has responded to this with the AI Act, the AI Regulation. It bans systems that categorise people biometrically according to characteristics such as ethnicity, religion, political conviction, or sexual orientation. Other categorisations remain permitted, but are considered high-risk applications. This means strict requirements: documentation, bias testing, human oversight.

From camera image to group assignment

Technically, it starts with a recording: an image, a video, a snippet of audio. A neural network – a program capable of learning that derives patterns from many examples – converts this recording into a long sequence of numbers. This sequence of numbers describes the characteristic in a form that a computer can compute with. A second part of the system then assigns the sequence of numbers to one of the predefined categories.

These categories are set by people, not by the machine. The model learns them from training data in which someone has previously labelled the examples. So whoever creates the labels decides which pigeonholes exist in the first place. And because categories such as gender do not fit into two boxes in reality, the system forces reality into too narrow a grid.

A typical misconception is that the result would be a fixed statement. In truth, the model outputs probabilities, for example 72 percent for a particular group. Only a threshold turns this into a decision. And the hit rates drop noticeably for groups that were rarely represented in the training data – for darker skin tones, this has repeatedly been measurable in studies.

Billboards, bouncers, and authorities

In everyday life, the technology is often embedded in advertising. Digital billboards in train stations estimate the age and gender of passers-by in order to adjust ads or count target audiences. Supermarkets and cigarette vending machines also use age estimation to replace ID checks.

In the news, the term usually appears in connection with surveillance. Reports about cameras that automatically flag members of minorities have strongly shaped the debate. Recruitment software that tried to derive personality profiles from video interviews also belongs to this field – it has since been largely discredited.

For companies, the topic is above all a legal question. Anyone offering a product with such functions in Europe must check whether it falls under the ban. When news reports mention fines against retailers or advertising networks, it is almost always about exactly this boundary.