Purpose: This study explores how university students use Large Language Models (LLMs) to personalize learning within a heutagogical framework, with particular attention to the opportunities and challenges that emerge in their learning experiences..Methodology: A qualitative exploratory design was employed involving 21 undergraduate students from science, social science, STEM, and education programs. Data were collected through semi-structured in-depth interviews, chat log analysis, and learning documentation, and analyzed using thematic analysis and trustworthiness criteria. Results: Four major themes emerged: LLM as a metacognitive scaffold, LLM as a learning partner, cognitive offloading, and autonomy versus dependency of the learning process. The findings show that while LLMs can stimulate reflection, critical questioning, and idea exploration, they also create the risk of shortcut thinking and reduced cognitive engagement. Conclusions: LLMs play a dual role in heutagogical learning, functioning as both cognitive catalysts and cognitive crutches. Their educational value depends on how they are designed and used to preserve productive struggles and learner autonomy. Limitations: This study was limited to a qualitative exploration of higher education and did not quantitatively measure learning outcomes. Contributions: This study contributes to the growing literature on AI in education by offering a nuanced understanding of LLMs as metacognitive scaffolds and informing pedagogical design for AI-scaffolded autonomy.