Reflections on the Impact Brought by Large Language Models to Software Engineering Maintenance Modes Under the Wave of Artificial Intelligence

Abstract

Software maintenance occurs throughout the life cycle of the system and is relatively expensive. The Introduction of large language models is changing how software maintenance is conducted. The five areas this paper investigates for the practical applications of large language models in software maintenance are: maintenance process, collaborative verification, legacy system transformation, maintenance cognition, and security governance. According to a survey of almost 5,000 technical professionals in 2025 by the DORA report, 90% of the practitioners have already used AI tools in their work, and more than 80% believe that productivity has increased. AI is only an "enhancer" of existing delivery modes; thus, only organisations with well-established platforms and processes will reap significant benefits, and otherwise, technical debt will accumulate at a high rate. According to the qualitative analysis of DORA, most of the time saved in the code generation stage is spent on review and verification. Therefore, the focus of maintenance work has shifted from writing to judgment, and the reliability of AI modifications is now more dependent on system-level understanding than on manual modification. Large Language Models have not solved the problem of maintenance; they have only moved it from the coding stage to the stages of verification, governance and architecture design.