Large language models (LLMs) such as ChatGPT are increasingly being used in development to generate program code, derive test cases, create software models, or link requirements to source code. In practice, textual requirements frequently serve as input for this purpose — however, these often contain ambiguities, inconsistencies, or contradictions. This can result in the AI writing unsuitable code, for example. Such linguistic weaknesses are referred to as "requirements smells." They have long been known in classical software engineering; however, their influence on AI-based tools has so far barely been studied systematically.

This is where the project "Requirements Smells in Prompts (ReSPro)" by Prof. Dr. Andreas Vogelsang and his team comes in. It conducts a fundamental analysis of how strongly the quality of language shapes the results of AI-assisted software development. "Many current AI systems work with requirements that are already difficult for humans to understand clearly, and these imprecise descriptions are then passed on to the AI systems," explains project lead Vogelsang. "For the first time, we will systematically analyze which types of ambiguities are particularly problematic for AI — and how developers can be supported in formulating better prompts."

To this end, the project examines various use cases, including automatic code and test case generation, model generation, and the tracing of requirements in source code. Based on the results, tools are also to be developed that automatically detect problematic phrasings in prompts and suggest concrete improvements or correct them directly.

The long-term goal is to make the use of AI systems in software development more reliable, more robust, and more transparent, thereby strengthening quality assurance in AI-assisted development processes.