Digital Signal Processing Using Generative Artificial Intelligence (DSgenAI)

Generative Artificial Intelligence (GenAI) opens new possibilities for automating complex tasks. For example, chatbots and AI agents have conquered the internet in recent years by replacing search engines, writing emails, or summarizing books. The underlying Large Language Models (LLMs) can efficiently process and reproduce natural language input as well as images and documents. Studies estimate that GenAI can unlock efficiency gains worth billions of euros in the economy [1].

Especially for safety-critical applications, a key prerequisite for any technology to achieve lasting practical efficiency gains is reliability and robustness. However, the statistical nature and hard-to-interpret learning processes of GenAI make this a real challenge, as there is fundamentally no guarantee of correct or helpful results (“hallucinations”). As a result, the gap between technologically possible and verifiably safe GenAI applications continues to grow [2].

A fundamental principle to making processes more transparent and verifiable is to combine LLM-based components with non-AI resources (see Figure 1). Our Fraunhofer IKS orchestrator framework decides, for individual process steps, whether the task should be solved by an LLM or not, and whether alternative plausibility-checking methods are available. Standards and established deterministic tools can serve as a decision-making basis.

Based on those principles, Fraunhofer IKS is developing various methods to ensure the safety of applications using GenAI, including:

• Improved retrieval-augmented generation (RAG) methods [3]
• Hybrid search methods
• Systematic review processes with human-in-the-loop
• Multi-modal plausibility concepts
• Uncertainty prediction
• Metrics for semantic similarity that provide certifiable safety margins
• Fine-tuning of private models for data-sensitive tasks

Benchmarks of workflows with realistic data show whether such approaches can meet the requirements of industry. Building on its unique expertise at the intersection between AI and safety, Fraunhofer IKS connects fundamental AI research with safety-critical industrial applications.

References

[1] McKinsey & Company, The economic potential of generative AI: The next productivity frontier, 2023. 

[2] A. Kreutz, C. Drabek, R. Beck, “Mind the Economic Safety Gap. Accelerating Safety Innovation using Generative Artificial Intelligence,” Embedded World Exhibition and Conference 2025.

[3] B. Balu, F. Geissler, F. Carella, J.-V. Zacchi, J. Jiru, N. Mata, R. Stolle, “Towards Automated Safety Requirements Derivation Using Agent-based RAG,” in Proceedings og the AAAI Spring Symposium, 2025, pp. 299–307.