(in German)

The article examines how Artificial Intelligence (AI) can improve risk assessment in occupational safety. Despite being a legal obligation, many companies, especially small and medium-sized ones, show deficiencies in their implementation. To address this gap, two AI-supported approaches are presented: First, using large language models (LLMs) and the Retrieval-Augmented Generation (RAG) method, systems can analyze unstructured text documents (laws, accident reports). They identify relevant hazards and suggest measures based on trusted, company-specific data. This approach reduces AI "hallucinations" and increases traceability. Second, inexpensive sensors can continuously monitor environmental data such as temperature or air quality. Through anomaly detection using neural networks (autoencoders), the AI can recognize subtle deviations from the normal state that are often overlooked during sporadic checks. Based on interviews and workshops, the authors derived 17 specific requirements for such systems, emphasizing transparency, traceability, and human decision-making authority (the human-in-the-loop principle). AI is understood as a "co-pilot" that supports experts, not replaces them. The final responsibility remains with the human. Therefore, the safe and trustworthy integration of AI requires explainability and careful design.

Please download the article "AI-support for processing textual and environmental data in risk assessments" (in German only).