Berlin’s violence detector is learning from synthetic fights and real pedestrians

Berlin’s behaviour-detection system at Kottbusser Tor must learn to distinguish violence from the ordinary movement of a crowded square. Walking, shopping and gesturing are abundant in public footage; labelled assaults are scarce. Mannheim addressed the same problem by staging 2,000 fights between police officers and estimates it needs 8,000 more. Berlin instead relies heavily on generated video.

The supplier Staige says some real movements were captured and abstracted, then a video generator produced variations with different people, lighting and times of day. Public examples contain obvious artifacts: figures appear not to touch, tram doors open while moving, and people smile, wave or hop unnaturally. The material may be privacy-preserving and cheap to scale, but it is not drawn from the distribution in which the detector must operate.

Data scientist Heiko Paulheim compares the problem to training a cat recognizer on children’s drawings and then testing it on photographs. Synthetic action scenes may overrepresent young men, dramatic motion, good lighting or racialized visual stereotypes. A detector trained that way can miss quiet forms of violence, such as taking a bag from an older person, while treating an innocent jogger in a hoodie as suspicious. False negatives and false positives arise from the same mismatch.

The system then learns from Kottbusser Tor itself. Police review alerts and label each scene as ‘alarm’ or ‘ignore’; Staige says those movement patterns are used to improve the model. Gait can function as a distinctive biometric signal, and Berlin’s 2025 police-law amendment permits training with personal data and transferring data to the software provider. The police’s public page says the current configuration phase does not store camera images, but it does not resolve how later feedback examples are retained, abstracted or shared.

This is a substantive follow-up to the scanner’s launch because it identifies the model’s epistemic dependency. ‘A human decides’ is only the final step; the system has already decided which moments the human sees. Before performance can be assessed, Berlin needs a documented test set representing the square, subgroup error rates, rules for retaining live examples and a clear account of how synthetic clips were generated. Without those, the trial evaluates an alert pipeline whose definition of violence is partly inherited from another AI.