Developer perceptions of LLM-assisted static analysis alert prioritization

Abstract

Static analysis tools surface code-quality issues but can produce more alerts than developers can address. This thesis investigates how professional developers perceive LLM-assisted alert prioritization. In a qualitative case study at Schuberg Philis, eleven developers compared three ranked SonarQube-derived alert lists: a baseline, an LLM ranking using static information, and an LLM ranking enriched with development activity and self-admitted technical debt. No condition was universally preferred. Developers judged actionability through priority logic, alert type, local context, and workflow constraints rather than tool metadata alone. The findings suggest that LLMs are most acceptable as bounded assistance on top of deterministic static analysis, with developers retaining approval, and that LLM-assisted static analysis workflows are already emerging in parts of professional practice.