The intensive use of generative AI tools in higher education raises questions about their impact on students' cognitive autonomy. This study compares students' interaction with these tools and examines factors potentially associated with differences in cognitive autonomy between students in the early semesters (2nd to 4th) and final semesters (8th to 10th) of STEM programs. Using a cross-sectional comparative mixed-methods design, cognitive autonomy is assessed with the self-regulation subscale of the Motivated Learning Strategies Questionnaire (MSLQ) and a prompt coding system measuring complexity, critical iteration, and cognitive delegation. Semi-structured interviews identified factors associated with the observed differences between groups. A sample of 188 STEM students participated (116 in early and 72 in final semesters). The results reveal distinct profiles of interaction with AI according to academic progress, with a delegative prompt profile predominating among early-semester students (75.9 %) compared to final-semester students (19.4 %), and a higher overall interaction-autonomy index among the latter (d = 3.92). These findings are consistent with an association, rather than a causal or developmental effect, given the cross-sectional design, between academic trajectory and a more critical, epistemically reflective use of this technology. The results indicate potential risks of technological dependence and a possible weakening of autonomous thinking among students with less disciplinary maturity.