In this study, interpretive phenomenology was adopted to explore how university students make sense of AI risks when using AI in language learning contexts. Eighteen undergraduate students were interviewed individually, with participants also completing experience journals during the two-week interview period. Three dialectical themes were found after analyzing all interview transcripts and journals: (1) "the panic of too much convenience." Students felt torn between the lure of being efficient and losing control, oscillating between saving time vs. taking shortcuts to learn, feeling empowered vs. dependent, affirming vs. losing authorship; (2) "hallucination" and "gatekeeper": AI's output had "hallucinations," which destroyed trust; students switched roles from producer to proofreader, experiencing physical anxiety about academic integrity; (3) "Is it still my learning?" cognitive offloading caused metacognitive doubt whether deep learning still took place once there were no longer any desirable difficulties. Our results suggest that risk does not belong to AI itself but rather is a meaningful world created by users under certain conditions. Instead of simply accepting/rejecting AI, students have developed context-based self-regulation strategies against AI risks. Our study tries to go beyond the limits of stage models for technology adoption and offers a phenomenological account of what it feels like to be a learner living in technologically mediated times. We argue here for turning the conflict into teaching moments so as to nurture critically minded AI literacies.