Detecting and Correcting Software Assurance Document

Large language models (LLMs) are a form of generative artificial intelligence (AI) that are increasingly being experimented with to improve different levels of efficiency in software engineering tasks. Assessing the limits of applicability of this fast-evolving technology is particularly important when considering the use of LLMs in high-stakes domains (such as healthcare, aerospace, and defense), where incorrect responses from LLMs can lead to catastrophic outcomes if not reviewed carefully by humans. These high-stakes domains also contain significant amounts of assurance-related activities, such as conformance to stringent requirements, demonstrating evidence of how the software meets these requirements, and documenting the relevant standards to ensure consistency. These activities are often captured in different documents, which introduce ample opportunities for errors resulting from inconsistencies. While generative AI is not deterministic, its summarization capabilities and its ability to generate information culled from vast data sources provide opportunities to detect software document incompleteness, inconsistencies, and discrepancies (DIID). This chapter explores LLM-based approaches that enable the detection and ultimately the correction of DIID. It reviews current approaches to DIID detection and demonstrates the use of LLMs through a drone avionics case study example.