Developer Perceptions on the Maintainability of AI-generated Code : A Survey-based Study

Sammanfattning

Background. AI-based coding assistants are widely adopted to accelerate routine tasks, yet their effect on long-term code maintainability and technical debt remains unsettled. Existing empirical work largely relies on objective metrics, leaving a gap in understanding how professional developers themselves perceive these long-term impacts. Objectives. This thesis investigates developer perceptions of AI-generated code maintainability compared to human-written code. Specifically, it identifies the practical challenges developers face and the strategies they apply to mitigate them. Methods. We conducted a cross-sectional, mixed-methods survey of professional software developers, guided by a rapid literature review. Closed-ended items provided descriptive quantitative frequencies, while open-ended responses were analysed thematically. Results. Based on 31 professional developer responses, the most frequent maintenance challenges are integration difficulties, technical debt accumulation, and hard-to-follow logic. Thematic analysis reveals that context blindness, scope drift, and comprehension burden are the dominant failure modes. To cope, developers heavily rely on persistent guardrails (reusable context and steering files), manual line-by-line review, and AI-assisted iteration. Conclusions. Developers manage the maintainability of AI-generated code not as an inherent property, but as a trade-off controlled by layered defences like human review and context-engineering. Furthermore, formal team-level governance of these tools still significantly lags behind individual developer practices.