<p>Contemporary efforts to innovate simulation-based maritime education increasingly involve the development of multimodal learning analytics (MMLA) systems that use large language models (LLMs) to generate automated feedback on student performance. These systems are often presented as solutions to challenges such as limited instructor capacity, inconsistent assessment practices, and demands for greater standardization and accountability in competence evaluation. While such developments create new possibilities for personalized feedback and enhanced instructor support, they also raise important questions concerning transparency, responsibility, professional judgment, and regulatory compliance in maritime education. This study examines how MMLA can be introduced within a policy landscape shaped by the Standards of Training, Certification and Watchkeeping for Seafarers (STCW), European higher education policies, and the European Union AI Act. Drawing on Cultural-Historical Activity Theory (CHAT), we analyze how maritime simulator instructors and policy experts (<i>n</i> = 12) position MMLA within the broader activity system of maritime education and training (MET), with particular attention to the relationships between new technologies, division of labor, and regulations. The analysis focuses on how MMLA reshapes instructor roles and feedback practices, as well as the contradictions and tensions that emerge between technological innovation, regulatory compliance, and the embodied, situated nature of professional judgment in simulator-based training. Our findings reveal tensions and contradictions between three fundamental aspects: innovation vs. regulation, automation vs. autonomy, and engineering design vs. pedagogical intent. The study concludes with implication for responsible implementation of MMLA in simulation-based maritime training.</p>