For decades, a camera had a relatively simple task: record. It collected images that someone might watch after an event, searching for a presence, a license plate or a gesture. Computer vision changes this sequence. A system can detect people and vehicles, follow a trajectory, count entries, recognize objects left in a place or signal a situation that departs from a configured rule. It is no longer only an archive of images; it becomes a filter that decides what deserves attention.
This can be useful when human attention is not enough. In a station, hospital, construction site or museum, software can help reduce hours of footage to inspect and send an alert about a door left open, a vehicle in a restricted area, a fall or unusual crowding. But ‘unusual’ is not a natural fact; it is a definition. Someone chooses which behaviors matter, which threshold triggers an alert and who receives the notification. The real question is not whether the camera sees better, but who wrote the rule through which it interprets what it sees.
This type of analysis needs to be distinguished from facial recognition. Detecting that a person crosses a space, or that an object remains unattended, is not the same as establishing the identity of the person in the image. Biometric recognition creates or compares a representation of the face and can lead to identifying, verifying or categorizing a person. These are different technologies with different consequences. Confusing them makes it harder to understand when a system is merely assisting an operator and when it is building a profile about someone.
Science fiction described this shift before it became ordinary. In *Minority Report*, screens do not merely observe; they anticipate and pursue. In *Blade Runner*, an image is enlarged until it becomes evidence to interrogate. *Gattaca* turns the body into an access key, while *Person of Interest* imagines a machine that decides every day which lives deserve attention. These works do not literally predict the present; they remind us that the problem begins when a technical gaze becomes a social decision.
The most important deeper issue therefore does not concern model accuracy alone. NIST tests show that recognition errors can vary by age, sex and demographic groups; image quality, capture conditions and the chosen threshold all affect the result. An alert should not be treated as guilt, nor a resemblance as an established identity. The more a technology is used to restrict access, stop a person or guide an intervention, the more human verification, decision logs and ways to challenge an error become necessary.
In Europe, the framework is not left to the goodwill of vendors alone. EDPB guidelines emphasize necessity, proportionality, transparency and data minimization in the use of video devices. The AI Act also introduces specific limits, including those concerning real-time remote biometric identification in publicly accessible spaces, with narrowly defined exceptions for law enforcement. This is not bureaucratic detail: it means a security project has to be designed as a technical, legal and organizational system at the same time.
For people who design spaces and images, this also opens a question of language. Can a place be safe without becoming opaque? Can a notice be understandable without revealing more data than necessary? Can a camera help prevent risk without turning every passerby into a suspicious variable? The quality of a system is measured not only by how many events it detects, but by how many false alarms it avoids, how visible it makes its own rules and how much control it returns to the people involved.
A GAZE OF CONTROL is not an invitation to demonize intelligent cameras. It is a reminder: when a device begins to select, classify and suggest, its vision enters the field of decision-making. AI can make security more timely; it cannot automate responsibility. That remains human, before the alert and after it.