Generative AI in Programming Education: An Empirical Analysis of Student Performance and Assessment Challenges in the LLM Era

Abstract

The integration of generative AI into computer science education has precipitated a paradigm shift, yet empirical evidence regarding its impact on student competency remains fragmented. This study presents a longitudinal analysis of 412 undergraduate students across three semesters, focusing on the critical "watershed"moment in early 2025 precipitated by the widespread adoption of high-performance domestic Chinese Large Language Models (LLMs) such as DeepSeek. Utilizing a mixed-method approach with questionnaires, code submissions, and examination data, we observed a structural transition in student behavior from sporadic tool assistance to infrastructure dependency. The quantitative results reveal a paradox: while LLM-assisted assignments demonstrated significantly higher accuracy (Mean Score Rate: 0.803) compared to self-written code (0.644), this performance advantage evaporated in closed-book examinations, where AI-reliant students showed no statistical improvement over their self-sufficient peers. Furthermore, error profile analysis indicates a migration of cognitive hurdles from basic syntax errors (Type S) to complex API misuse (Type A) and logical hallucinations (Type L). Notably, despite the automation of coding tasks, temporal efficiency did not significantly improve, suggesting that the cognitive load has merely shifted from code construction to verification and debugging. In light of these findings, particularly the evidence of "competency hollowing", this paper proposes a pedagogical pivot towards process-oriented assessment, adversarial learning tasks, and a 3-layer curriculum structure to ensure that AI serves as a scaffold for, rather than a substitute of, fundamental computational thinking.