Disputed matching of protest footage by Baden-Württemberg police and Berlin's behaviour-recognition cameras

An officer of the State Criminal Police Office (LKA) of Baden-Württemberg in Germany testified in a criminal case before the Stuttgart Regional Court. netzpolitik.org's report of 21 October drew on a Stuttgart weekly's coverage of the hearing and on statements from a defence lawyer for the accused.

The officer said that, to check whether two of the accused had already come to police attention, he had analysed footage of a registered demonstration in Ulm. Besides the two accused, the matching identified four other participants who, to date, have no connection to the case. The report's headline and subheading refer to facial recognition, but the officer's testimony did not say which system was used. He said he had followed a "principle of completeness" and checked whom else the computer could recognise, and the report gives no source for this principle.

According to the report, the witness did not answer in court who ordered the matching or on what legal basis. In its reply to netzpolitik.org, the LKA stated that it could not comment on ongoing proceedings and that personal data is collected in line with the relevant legal provisions, and it did not respond to the question about the "principle of completeness". A lawyer at the German civil rights group Gesellschaft für Freiheitsrechte (GFF, Society for Civil Rights) argued that no such principle allows the police to identify everyone who has ever demonstrated alongside a suspect. Without concrete indications that these people were accomplices, the lawyer considers such investigative measures unlawful.

The report does not name the system the officer used. Its background section covers the facial recognition system of Germany's Federal Criminal Police Office (BKA), with no statement that it was the one used in this case. The system has run since 2008 and compares surveillance camera images or wanted-person photos with photos in INPOL, the police database. Figures cited in the same report put INPOL at about 7.6 million photos of some 5.4 million people in January 2026, with federal and state criminal police offices running about 313,500 searches in 2025.

Since 6 October, Berlin police have been testing a different kind of camera at Kottbusser Tor, an area the city's interior department lists as a high-crime location. According to the police explainer page, the system recognises situations and does not analyse faces or personal characteristics. Administrative rules published in October by the Senate Department for the Interior and Sport (Senatsverwaltung für Inneres und Sport) give beating, kicking and people lying on the ground or falling as example situations. During the test phase the system neither detects nor raises alerts, but footage is stored.

According to the department's answer in September, the round-the-clock phase is expected to start in the fourth quarter of 2026.

Under Berlin's General Security and Public Order Act (ASOG), amended in December 2025, automated remote biometric identification on footage from the high-crime-location cameras is ruled out, as is using that footage to test or train remote biometric identification programs. Remote biometric identification means recognising a person from a distance, without their involvement, by matching features such as the face. Under the same amendment, police may separately, with a court order, match face and voice data of specific people after the fact against personal data publicly available on the internet.

According to the police page, the system is switched off during assemblies whenever participants might be caught on camera. Footage from the test phase onwards is kept for 28 days.

Perspective

Real-time identification matches and recognises people while they are being filmed, and post-event identification runs the matching on footage already recorded. The Ulm case involved post-event matching, and according to official descriptions Berlin's system does neither and only assesses what is happening in the frame.

The EU's Artificial Intelligence Act (AI Act) has different rules for the two. Under Article 5, in force since February 2025, real-time identification by law enforcement in publicly accessible spaces is prohibited, with three exceptions.

The three are searching for victims of abduction, trafficking in human beings and sexual exploitation and for missing persons, preventing a specific and imminent threat to life or a terrorist attack, and locating suspects of serious crimes. Each must be strictly necessary, and member states must expressly allow it in national law. Each use requires prior authorisation from a court or an independent authority, and in urgent cases the request must follow within 24 hours at the latest.

Post-event identification falls outside the prohibition and is classed as high-risk to the extent its use is permitted under EU or national law (going by the wording of Annex III, our reading is that the AI Act itself does not authorise such use). Under Article 26(10), an investigation targeting a specific suspect requires an authorisation request in advance or within 48 hours at the latest, except for the initial identification of a potential suspect based on facts directly linked to the offence. Untargeted use with no link to a criminal offence, criminal proceedings, a genuine threat or the search for a specific missing person is also banned.

These high-risk rules were originally due to apply in August 2026, and under a simplification amendment (the AI Omnibus) in force since July 2026 they now apply from December 2027.

The Ulm case involved post-event matching on footage already recorded, but with the system unknown, it cannot be confirmed whether it counts as remote biometric identification as the AI Act defines it. The Act's authorisation rules did not yet apply at the time, so legality turns on German national law, which is exactly what GFF questions. According to the police page, Berlin's system recognises situations, not identities. In June, Berlin police's chief digital officer told a parliamentary committee that, by the force's own assessment, the system would probably count as low-risk under the AI Act.

The European Commission's 2025 guidelines on prohibited practices include an example in which police pull footage from street CCTV and use post-event facial recognition to match it against social media photos to find the organiser of a protest. The guidelines' assessment is that this use is not prohibited and counts as high-risk. According to a footnote, it is questionable whether post-event identification is strictly necessary to identify the protester.

The example is drawn from a judgment of the European Court of Human Rights. In that case the court held that, with no risk to public order or transport safety, both post-event and live use of facial recognition were disproportionate. The guidelines themselves are not legally binding.

Identification and anonymisation are one point of contention in Berlin. In an interview on 22 September, Berlin's data protection commissioner said that the law rules out automated biometric identification but does not prevent a specific person being identified manually, case by case, from stored footage. Manually building profiles of movements and social contacts within the monitored area also remains technically possible.

At a hearing of the Berlin parliament's interior committee in June, an expert from the German hacker collective Chaos Computer Club (CCC) argued that stick-figure images do not count as anonymisation, since timestamps alone link them back to the original footage. At the same session, the senator responsible for the interior (Innensenator) said that the system carries out no biometric or other automated matching of personal data. The senator also said it would not be expanded into citywide surveillance, and according to the administrative rules, the vector data the system produces cannot be reconstructed into images.

Effectiveness is the other point of contention, and in its September answer the interior department sets the system's goal as early detection of violence and of people on the ground who need help. A researcher who studied the Hamburg pilot told the June hearing that situations police officers mentioned in interviews as triggering alerts included hugs, children kicking footballs, people pushing prams and people lying on the ground (Hamburg uses software developed by a different institution). The researcher said that researchers have no precise information on how many false alarms occur at each site, or what they are. Hamburg police have published one figure, which the researcher considers unreliable.

As of its 6 September answer, the city government had no plans to commission an external academic evaluation.

Berlin held a state election on 20 September, and according to netzpolitik.org, the exploratory paper of the SPD, the Greens and the Left Party (the preliminary document drawn up before coalition talks) includes an evaluation of video surveillance in 2028. The three parties can begin negotiations only after the Left Party votes at its party conference on 31 October. In August, the Greens' and the Left Party's parliamentary groups asked Berlin's Constitutional Court to review the police law, including the behaviour-recognition cameras. They did not seek a suspension, so the law remains in force until a ruling.

In a coalition agreement published in May 2026, the Greens and the CDU in Baden-Württemberg pledged to expand the Mannheim model, on the grounds that AI video surveillance can serve both security and fundamental rights. Mannheim has piloted behaviour-recognition cameras for hitting, kicking and similar acts since 2018, and the agreement includes adding remote biometric identification at the pilot sites, without saying whether real-time or post-event. According to netzpolitik.org's report in May, this would require amending the state police law. As of 24 October, we had found no record of such an amendment.

According to the same May report, Frankfurt was at the time the only place in Germany using real-time facial recognition on public cameras, and in April netzpolitik.org reported that each search needs a judge's order.

People who spend time near Berlin's pilot sites can tell from signs where these cameras are. Under ASOG, police must put up signs and publish the locations, and the administrative rules list four pilot sites. The cameras are switched off during assemblies, but footage of people passing at other times is kept for 28 days and may be viewed manually during that period. For people who join protests, even away from Berlin's pilot sites, post-event matching of existing footage like that in the Ulm case is currently governed by German national law.

Under Article 9 of Taiwan's Police Power Exercise Act, when facts give sufficient grounds to believe that the conduct of participants in an assembly, parade or other public event may endanger public safety or order, police may photograph, record or use other technical means to collect participants' on-site activity data during the event, extending to third parties where unavoidable. The police make that judgement on the facts, and the article does not require court authorisation. Data must be destroyed when the event ends unless it needs to be kept to investigate a crime or other violation, and retained data must be destroyed within one year of its creation at the latest, except in cases indicted and still on trial or concerning violations of the Organized Crime Prevention Act.

We, a community based in Taiwan, found no law that expressly regulates police use of facial recognition matching. Neither the Police Power Exercise Act nor the Personal Data Protection Act, the two laws we searched, contains any reference to faces or biometric features, and we did not read the National Police Agency's internal rules.

In Japan, police register photos of arrested suspects in a database managed by the National Police Agency, following the National Public Safety Commission's Rules on the Management and Use of Suspect Photographs, under which the photos are kept to assist criminal investigations; the rules make no mention of facial recognition or matching. Answering questions in the Cabinet Committees of the House of Representatives in March 2021 and the House of Councillors in May 2021, police officials explained that when police obtain a suspect's facial image from security cameras in an individual case, for example with the cooperation of the camera's operator under the Code of Criminal Procedure, they match it against suspect photos held by the police to help identify the suspect. In the March answer, the official put the police's suspect photo data at about 11.7 million items at the end of 2020, including people whose acquittals had become final, and said keeping those photos posed no legal problem, while photos from wrongful arrests are deleted. We found no law or National Public Safety Commission rule that expressly regulates this matching, and we did not look into rules on police filming at assemblies in Japan.

In South Korea, we found no statute specifically on police filming at assemblies, and the practice rests on the National Police Agency's internal Rules on Evidence-Gathering Activities at Assemblies. Under the rules, filming for criminal investigation is generally limited to while or just after a violent or other crime is committed (or shortly before, when harm to others is imminent and evidence urgently needs to be secured), to the minimum extent necessary, and the people being filmed must be notified in advance. Footage not needed for an investigation must be deleted as soon as the assembly ends, footage may be entered into an evidence-reading program to confirm suspects' identities, and footage of unidentified people still needed for an investigation may be kept until the limitation period for the offence expires.

Under the Framework Act on the Development of Artificial Intelligence and the Establishment of Trust, in force since 22 January 2026, AI systems that may significantly affect life, bodily safety or fundamental rights and are used to analyse biometric data such as faces for criminal investigations or arrests are classed as high-impact AI. AI businesses that provide them owe risk management, explanation, user protection, human oversight and record-keeping duties, and the act defines AI businesses as those that develop and provide AI or provide products or services using it, state agencies included. We found no dedicated law or police guideline on facial recognition matching in South Korea, and we did not check whether police have applied the new act in practice, or whether police using such a system count as an AI business under the act.

Readers can compare rules across places with a few questions. First, whether the law covers real-time identification, post-event identification or both. Second, whether matching needs prior authorisation from a court or an independent authority, and how data on people unconnected to a case who appear on camera is handled. Third, whether governments publish the number of uses and error rates so that outsiders can assess how well the systems work.