The Baden-Württemberg coalition agreement, signed by the Greens and CDU on 11 May 2026, commits the state to becoming Germany's first to run a public-space surveillance system that combines live facial recognition with automated behavior analysis. The pilot is to launch in three locations — Mannheim plus two cities not yet named — and the deal expressly anticipates biometrische Fernidentifizierung and the matching of camera feeds against publicly available images on the internet.
That last sentence is doing a lot of work. Mannheim has run behavior-analysis cameras since 2018; the system flags movement patterns like hitting, kicking or running and was sold to residents with an explicit municipal promise that biometric facial recognition would not be used. Eight years on, the coalition is proposing to add precisely the feature the city said it would not add — and to go further, by mining public internet photographs to expand the matching base. The expansion requires amending the state's Polizeigesetz, which has not yet been introduced, let alone passed. The State Commissioner for Data Protection, Prof. Tobias Keber, has issued a public statement noting that the agreement would introduce "extensive digital surveillance powers" restricting fundamental rights, and that any such tools must be controllable, trustworthy and proportionate.
The case for it, fairly stated
Mannheim is not a hypothetical. The Alter Messplatz and parts of the Innenstadt have seen real violent-crime clusters, and the existing behavior-analysis system has, on the city's own account, helped operators triage incidents in real time without identifying individuals. The coalition's framing is also blunt: KI-Videoschutz kann für mehr Sicherheit bei gleichzeitiger Schonung der Grundrechte sorgen — AI video protection can deliver more security while sparing fundamental rights. The state is signalling, defensibly, that AI tooling for law enforcement should be regulated and used, not categorically banned. A pro-innovation policy has to grapple with the reality that surveillance technology is being deployed in democracies with constitutional courts, judicial oversight and freedom-of-information regulators — not just in autocracies.
Where the plan meets EU law
The problem is not behavior analysis. The problem is what the coalition has bolted on top.
Article 5(1)(h) of the EU AI Act, in force since 2 February 2025, prohibits the use of "real-time remote biometric identification" in publicly accessible spaces for law-enforcement purposes. The Regulation carves out three narrow exceptions — searching for trafficking or abduction victims and missing persons; preventing a specific and imminent terrorist or life-safety threat; and locating a suspect in a serious crime carrying at least a four-year custodial floor — and conditions each on a Fundamental Rights Impact Assessment, prior authorisation by a judicial or independent administrative authority, registration in the EU database under Article 49, and notification to both market-surveillance and data-protection regulators. A standing city-centre deployment is none of those things; it is exactly the everyone-is-scanned-just-in-case architecture the Act was drafted to outlaw.
The coalition's second ambition — matching camera feeds against publicly available images on the internet — runs into a separate AI Act prohibition. Article 5(1)(e) bars the "untargeted scraping of facial images from the internet or CCTV footage" to build or expand recognition databases. This is the rule that drained Clearview AI out of Europe: the Dutch Autoriteit Persoonsgegevens imposed a €30.5 million fine in May 2024, on top of €20 million penalties each from France's CNIL, Italy's Garante and Greece's HDPA. Every major European DPA that has examined the Clearview model has held that scraping public web photos to power facial-search functionality is unlawful biometric processing. A state government cannot procure, as a service, what a private company is forbidden from offering.
A proportionate alternative exists
None of this requires Baden-Württemberg to scrap the Mannheim model. The behavior-analysis layer — pattern detection without identification — can survive as a triage assistant, and the state has a respectable case for tightening its statutory basis and oversight rather than abandoning it. What the coalition cannot do, without colliding with the law it claims to respect, is graft real-time facial identification and internet-image scraping onto that base.
A proportionate-regulation path would: (1) keep the existing system as an assistance tool with an explicit statutory ban on identifying individuals; (2) make any future remote biometric identification case-by-case rather than ambient, with judicial pre-authorisation, FRIA publication, and a serious-crime trigger that mirrors Annex II of the AI Act; (3) explicitly prohibit the use of internet-scraped face databases in the amended Polizeigesetz; and (4) sunset every camera deployment after a fixed period unless re-reviewed by the LfDI. None of this stops the police from using AI; it stops the state from running an always-on biometric checkpoint in three cities.
The coalition agreement is, in that sense, less a security policy than a test case. If the amended Polizeigesetz is enacted and survives review by Keber's office, the Bundesverfassungsgericht, and ultimately the Court of Justice of the EU, the contours of Article 5 will have been drawn by the deployment rather than by the legislation. That is the wrong way around — and it would give the rest of Europe an unhelpful precedent to copy.