Large Language Models (LLMs) offer scalable opportunities to personalize feedback in education, yet their trustworthiness and effectiveness remain underexplored. We present a study conducted in an introductory programming and data science course with approximately 1,400 first-year university students. A subset of these students received both peer and LLM-generated feedback on their individual programming projects. Our results show that 56% of students preferred the LLM feedback, and 52% could not reliably distinguish it from human-written feedback. Student ratings suggest that LLM feedback is perceived as helpful, constructive, and relevant, though it often lacks personalized depth and motivational nuance. These findings underline the potential of LLMs to support scalable, personalized education, while pointing to key areas for responsible improvement. Based on these insights, we outline the future roadmap for the course in which LLM-generated feedback supports students in their learning journey but also instructors through monitoring student performance and helping to allocate instructional resources more effectively. Given limited human resources this approach enables personalized instructor feedback to be scaled to a large group of students.