The ongoing digitalization of public administration means that large language models (LLMs) are increasingly being integrated into everyday work. However, this development also raises the bar for digital sovereignty. In particular, this raises the question of how the federal administration can avoid technological dependencies while at the same time using powerful AI systems securely.

The current study “Digital Sovereignty and Large Language Models in the Federal Administration” analyzes the role of large language models in the federal government and evaluates existing in-house developments with regard to three key sovereignty goals: changeability, design flexibility, and influence on providers. The study was based on comprehensive data collection on ongoing AI projects and an international market and measures analysis.

The results clearly show that, thanks to AI systems developed in-house, the administration does not have to rely on proprietary products from large, primarily non-European technology companies for many typical LLM use cases. This reduces the risk of entering into new dependencies or deepening existing ones. In addition, the solutions currently in use are primarily used as work support – a failure would therefore not jeopardize the administration's ability to act. Interdepartmental systems in particular are already being developed in such a way that digital sovereignty criteria are consistently taken into account.

Ensuring digital sovereignty in a modern administration

The study also shows that LLMs are mainly based on non-European open-source solutions hosted on internal administrative hardware. In view of changing open-source conditions, the authors recommend examining the development of a European LLM that better reflects administrative requirements and ensures long-term independence. At the same time, the market analysis raises questions about the development of a broader ecosystem for government LLM applications – for example, through promotion, targeted procurement, or federal cooperation such as in the Germany Stack. However, the study also points to challenges: complex legal requirements, uncertainties in provision for reuse, and a lack of specialized AI infrastructure and skilled personnel. In a series of concrete recommendations for action, the study therefore calls for, among other things, shared infrastructures, clear legal guidelines, stronger open-source initiatives, and additional training measures.

The conclusion: The federal government's current LLM initiatives are already strengthening digital sovereignty and offering functional alternatives to external solutions. However, targeted further developments are necessary to remain independent and capable of action in the long term.