The rapid integration of large language models (LLMs) in higher education presents significant opportunities for personalized learning alongside profound ethical challenges, such as academic dishonesty and cognitive over-reliance. Moving beyond technology acceptance models, this study investigates the structural drivers of ethical LLM usage among Chinese university students by integrating social cognitive theory (SCT) and rational choice theory (RCT). Utilizing an adapted, culturally validated multidimensional scale, survey data from 519 students were analyzed using confirmatory factor analysis (CFA) and partial least squares structural equation modeling (PLS-SEM). The findings reveal a stark contrast between institutional and interpersonal influences: while abstract academic ethical climates and mere policy awareness fail to ensure ethical behavior, direct ethical supervision shows a strong positive association with students’ moral consciousness and ethical use (SCT). Conversely, prolonged use of LLMs is negatively associated with ethical behavior, suggesting that as students gain technical proficiency, their rational cost-benefit calculations often tip toward unethical shortcuts unless credible deterrents—such as detection probability and sanction severity—are present (RCT). This study contributes to the theoretical understanding of AI ethics by demonstrating that ethical technology use is a negotiated outcome between social modeling and rational risk assessment. Practically, it urges institutions to transition from passive policy publishing to active supervisor training, balanced oversight, and the cultivation of peer-level moral self-efficacy.