Berlin's behaviour scanner is now learning from everyone who passes through Kottbusser Tor
Berlin activated AI-assisted video surveillance at Kottbusser Tor on 28 October. Thirty cameras cover the square, while software supplied by video-AI company Staige searches the live stream for movements associated with fighting: punches, kicks, violent pushing or pulling, and struggles on the ground. Human operators review alerts before police are dispatched.
The system was trained with real and synthetic scenes. According to the vendor's representative, it will also learn from the place it watches: when an operator confirms or rejects an alert, movement patterns captured at Kottbusser Tor can become additional training material. That turns a public square into a continuing data source for a commercial surveillance product, not merely the location where a finished tool is used.
Evidence that the underlying method improves policing remains thin. Netzpolitik reports that Mannheim cannot identify a single crime prevented or solved after eight years of training a comparable scanner. In Hamburg's trial, the limited figures released by the state implied 99 false alarms for each useful alert. Berlin is betting that a different vendor, synthetic training data and local feedback can overcome that base-rate problem.
The legal and political sequence is equally important. Berlin's outgoing CDU-SPD coalition created an exception allowing surveillance data to be used for training and shared with manufacturers. The SPD-led interior administration then activated the system while the Greens and Left—now negotiating a government with the SPD—challenge the expanded police powers in court. Cameras are planned for further sites including Warschauer Brücke, Alexanderplatz and Görlitzer Park.
This is the operational sequel to earlier procurement reporting: the behaviour scanner is no longer a specification in leaked documents. It is producing alerts, operator judgements and training examples from a neighbourhood where homelessness, addiction, commerce and ordinary conflict overlap. Because residents cannot know what motion triggered an alarm or how to behave so as not to be flagged, the system shifts police suspicion from an observable act to an invisible statistical interpretation.