AI in public authorities: where the state already uses artificial intelligence and where the legal limits lie

AI has arrived in day-to-day government

For a long time German administration was seen as a laggard in digitalisation. That is changing. In the federal government's portfolio alone, dozens of AI use cases are documented—from pattern recognition in infectious diseases and evaluation of remote sensing data to combating VAT fraud.

Unlike a digital corporation, the state may not freely analyse its data holdings but only within its statutory mandate and clearly defined competences. Under Federal Constitutional Court case law it may not create personality profiles of citizens.

Where security authorities use AI today

The most practically important field is internal security. Several federal states have worked for years with location-based prognosis tools to identify possible crime hotspots from large datasets and deploy police more targeted. This so-called predictive policing is meant to direct scarce resources where statistics suggest offences are likely.

Investigative authorities also evaluate social media (social media monitoring) and merge information from different sources. Merging datasets and automated analysis has become a legal focal point.

Further reference cases of automated administration

Automated data analysis is only one building block. Constitutionally settled or intensively discussed areas include automated licence plate recognition, passenger data matching under the Passenger Data Act, and anti–money laundering monitoring. Video surveillance with algorithmic pattern recognition is also increasing—for example detecting hazardous movement patterns or biometric facial recognition. Each area has its own fundamental-right sensitivity and legal basis.

Service not intervention: chatbots and language models in administration

Not every AI use is a fundamental-right intervention. A growing share serves citizen service. Authorities deploy chatbots that provide information regardless of opening hours—for elections or short-time work benefits applications. Baden-Württemberg is testing an internal assistant with language model F13 to summarise study results and prepare notes. Such applications raise less intervention law than data protection, copyright, and organisational questions—e.g. reliability of outputs and allocation of responsibility.

What this means in practice

Some guidelines emerge that every authority and technology provider in the public sector should heed. Benefit of AI alone does not justify intervention. What matters is statutory authorisation. The deeper a system intervenes in fundamental rights, the higher the threshold and the protected legal interest must be. The legal basis must regulate type and scope of data and permitted analysis methods with sufficient precision. Democratic legitimacy and clear allocation of responsibility are not formalities but preconditions for sovereign decisions to be supported by AI at all.

Anyone introducing, procuring, or legally accompanying an AI system in the public sector should examine these requirements from the design stage onward—not only at the end.