The integration of Generative Artificial Intelligence (GenAI) into engineering education raises a critical tension between operational efficiency and the development of deep cognitive competencies. This study examines this dichotomy in 162 Civil Engineering in Computer Science students at a Chilean state university. Using a quantitative descriptive–correlational design, it investigates the relationship between usage patterns and perceptions of academic self-efficacy. The results reveal a “paradox of efficiency”: while 86.4% of students use the tool as conceptual support and 72% value the acceleration of code debugging, 60.9% acknowledge a significant risk of cognitive dependency that threatens their problem-solving autonomy. The analysis suggests that, in contexts with limited personalized tutoring, AI assumes the role of a “Shadow Tutor,” boosting short-term performance while generating ethical and epistemic friction in the long term. The study concludes by highlighting the need to transition toward a “pedagogy of verification,” in which assessment focuses on the critical auditing of AI-generated solutions rather than on syntactic coding alone.