Berlin behavior scanners: police want to train AI on real data

Berlin is advancing AI-assisted video surveillance. On 7 September 2026, netzpolitik.org reported that the specification for the "project to test AI-assisted video technology" had been made public. This happened during an IT attack on the Berlin administration, in which the authors say around 1.5 million files were leaked.

What the behavior scanners are supposed to do

The plan covers fixed and mobile camera systems as well as drones. The system captures everything moving around protected objects. It triggers an alarm when objects or movement patterns stand out. A police officer then checks the hit. The software does not decide alone, but it pre-sorts the results.

The conflict lies in this pre-selection. Behavior scanners use features captured from a distance to produce hit lists for the people who make the final call. Error rates and systematic biases are not a downstream operational issue. They are part of the technology itself.

New law, new data paths

The second element followed a week later. On 15 September 2026, netzpolitik.org reported on the amendment to the police law. It allows personal data to be processed further to test or train AI systems. A minor interpellation by the Left party asked whether non-anonymized data flows to private companies. The answer states that effectively or irreversibly anonymized data sets are transmitted, unless anonymization prevents the purposes of testing or training. The report names Staige GmbH, part of the Adesso SE environment, as an example of the companies involved.

The classification

This shifts the question. The focus is no longer on the testing itself, but on the data basis. Images from public space that accumulate during operation can become training material. Whoever creates the surveillance archives may later decide how much of it lives on in models.

What remains open is the assessment. How often and in which situations the system actually triggers an alarm. Whether the cameras hang only at protected objects such as embassies or courts or also capture open squares. And which deletion periods apply to training data. These figures are essential for the debate about proportionality. Without them, the assessment of the technology stays with the declarations of intent in the specification.

The police officer decides in the end, the system selects beforehand, and in future may learn from images from exactly this operation.