Generative AI introduces a fundamental structural mismatch into software engineering: code is produced through stateless probabilistic generation, yet integrated into systems that evolve through stateful, temporally constrained architectures. While existing discourse concentrates on hallucinations, isolated defects, and productivity gains, it lacks a unified account of the degradation mechanisms unique to AI-assisted development—mechanisms that are neither reducible to individual bugs nor adequately captured by traditional notions of technical debt. This paper proposes a three-layer taxonomy of structural degradation in AI-generated code, organized across the Intent, Structural, and Behavioral layers. We first establish the root cause model: a three-dimensional discontinuity—temporal, authorial, and state-based—arising from the interaction between stateless generation and stateful system evolution. We then define and differentiate eighteen distinct degradation phenomena distributed across the three layers, each characterized by its definition, distinction from adjacent concepts, illustrative examples, and degradation path logic. Finally, we demonstrate that these phenomena propagate across layers through identifiable causal mechanisms, and articulate the Visibility Inversion Principle: detection probability is inversely proportional to origination layer depth, while damage magnitude is directly proportional to failure layer depth. The contribution of this paper is foundational. It establishes a coherent ontology of degradation patterns and their cross-layer propagation dynamics, providing a necessary conceptual foundation for future analytical, governance, and architectural research in AI-assisted software engineering.
Spark Tsai (Mon,) studied this question.