The Impact of AI-Assisted Debugging on Software Maintenance Efficiency

Thesis Abstract

The rapid adoption of AI-assisted debugging tools in software engineering promises to transform maintenance efficiency by accelerating fault localization, reducing debugging time, and improving code quality; however, empirical evidence on their real-world impact across diverse project contexts remains limited. This study addresses the problem by evaluating how AI-driven debugging supports influence maintenance outcomes in industrial and open-source settings, with particular attention to accuracy of fault localization, time-to-resolution, and post-fix defect density. The aim is to quantify the effect of AI-assisted debugging on maintenance efficiency and to identify contextual factors that modulate its effectiveness. The specific objectives are (1) to measure changes in time-to-fix and defect leakage before and after integrating AI-assisted debugging across 12 heterogeneous software projects; (2) to assess the accuracy of AI-powered fault localization compared with human expert debugging using a matched-pairs evaluation on 180 debugging episodes; (3) to examine the impact on code churn and maintainability metrics over a six-month post-fix window; (4) to explore developers’ perceptions of AI-assisted tools and the organizational factors that influence adoption, using thematic analysis of semi-structured interviews with 30 practitioners; and (5) to test a theoretical model grounded in the Technology Acceptance Model and the Theory of Cognitive Load to explain variance in maintenance efficiency outcomes. The methodology adopts a mixed-methods research design that combines quantitative performance metrics with qualitative insights. The population comprises software maintenance teams from five mid-to-large organizations and two open-source communities, encompassing a total of 22 projects. A stratified purposive sample yields 12 projects for quantitative measurement (with historical data spanning 24 months pre- and 12 months post-adoption) and 10 projects for qualitative inquiry. Data collection instruments include project issue trackers (for time-to-fix, re-open rates, and post-fix defect density), version control metadata, standardized fault localization accuracy benchmarks, and a structured survey instrument measuring perceived usefulness, ease of use, and cognitive load. Additionally, semi-structured interviews are conducted with developers and team leads to capture adoption dynamics and workflow integration. Validity and reliability are ensured through triangulation, pilot testing of instruments, and inter-rater reliability checks on defect categorization. Quantitative data are analyzed using difference-in-differences regression to estimate the effect of AI-assisted debugging on maintenance efficiency, complemented by multilevel modeling to account for project- and team-level heterogeneity; fault localization accuracy is evaluated via ROC analysis and paired t-tests, while time-to-fix and defect density are analyzed with time-series interventions. Qualitative data undergo thematic analysis following Braun and Clarke’s approach, with coding performed by two researchers to ensure credibility, and a joint display synthesizes quantitative and qualitative findings. The study expects to find that AI-assisted debugging reduces mean time-to-fix by 25–40% across projects, with a corresponding 10–25% reduction in post-fix defect density, and improved fault localization accuracy (AUC improvement of 0.08–0.12). However, effectiveness is anticipated to be moderated by project complexity, tool integration maturity, and developer cognitive load, with higher gains in projects adopting automated test generation and continuous integration pipelines. The contribution to knowledge includes (i) robust empirical evidence on the impact of AI-enabled debugging on software maintenance efficiency across diverse environments, (ii) a validated analytical model linking technology acceptance, cognitive load, and maintenance outcomes, and (iii) practical guidelines for heterogeneous teams on selecting, integrating, and governing AI-assisted debugging tools. The researchers conclude that AI-assisted debugging offers meaningful efficiency gains when accompanied by alignment with existing development workflows, targeted training, and careful management of cognitive demands on practitioners. Recommendations include investing in tool interoperability, developing domain-specific fault patterns, establishing governance for AI-assisted recommendations, and conducting ongoing post-deployment impact assessments to monitor effectiveness and maintainability.