Arrest warrant rejected: AI hit is only a vague hint for judges A district court slows down the use of facial recognition and strengthens the rights of defendants against opaque IT investigation tools. In the world of law enforcement, automated successes in manhunts sound efficient: an image, a database match, a hit. However, what is technically possible does not always withstand legal scrutiny. The Reutlingen District Court made it clear in a decision published on February 11: Algorithmically generated identification clues are not sufficient without sound substantiation and technical transparency to put someone behind bars. The proceedings originated from an incident in a drugstore in October 2025. Employees observed via video surveillance how a person stole several bottles of women's perfume. When the suspect entered the store again a short time later and was approached, it escalated: fleeing, the perpetrator swung around with an umbrella and hit two employees who were trying to hold him. The police used the video material for a facial recognition search at the Federal Criminal Police Office (BKA). The system delivered a hit: a man known to the police, who was already wanted for other offenses. Based on this “match” and the blanket classification as having a prior record, the public prosecutor's office applied for an arrest warrant for robbery. “Ominous” Software: Reprimand lacking Transparency The Reutlingen District Court ( 5 Gs 19/26). The reasoning attacks the current practice of AI-supported investigations. The judges described the facial recognition software used as downright “ominous.” The accusation: Neither the functionality nor the algorithm, the reference data used, nor the error rates were comprehensibly documented. The BKA operates the official police facial recognition system (GES). Last year, German authorities used the technology significantly more often for identifying people than before. With a total of around 343,856 searches in