mindUp at the Digital Day 2025 of the Freiburg Regional Council
The sensitive handling of data in AI systems was the subject of a presentation by mindUp Managing Director Ralf Walther. AI in an administrative context triggers specific requirements that must and can be explicitly fulfilled, as became clear in the course of the presentation.
Principle of equality
Compliance with the principle of equality is one of the fundamental requirements for the use of AI in administration. A major challenge is that AI models can make discriminatory decisions due to biases in the training data or systematic inequalities. To prevent this, targeted fine-tuning of the AI systems and the use of subject-specific large language models (LLMs) are recommended to ensure the fairest and most balanced decision-making possible.
Quality & verifiability
AI-based decisions must be of high quality and verifiable. The challenge here is that LLMs can sometimes generate incorrect information (hallucinations) or provide inconsistent output. By introducing organizational and technical correction and verification processes, the quality of the results can be ensured and permanently checked.
Transparency and legal certainty
Transparency and legal certainty are of central importance for administrative action, especially in the case of decisions based on AI. A key problem is that LLMs often do not have the necessary expertise in legal or administrative topics. In addition, the often creative expression of AI models in legal language is problematic. Strategies for the comprehensibility of the results, the use of explainable AI and a more deterministic design of the results with reduced creativity provide a remedy.
Data protection and data sovereignty
In connection with the use of external AI services, the focus is on protecting personal data and ensuring data sovereignty. Challenges arise, among other things, from the use of external providers and a possible lack of knowledge on the part of employees with regard to data protection law and the GDPR. These risks can be effectively countered by implementing general data protection measures, developing a comprehensive data strategy and providing employees with targeted training and qualifications in data literacy.
Reliability
Another key concern is the reliability of the AI systems used. Especially in the case of external AI services, updates or further developments of the LLMs can change functionalities and behaviors. To ensure reliability in the long term, careful versioning of the AI components used and regular compatibility tests for updates are essential. This ensures a stable and predictable system landscape.