Researchers at Sungkyunkwan University have published a comprehensive study examining the dual impact of generative artificial intelligence on educational outcomes, revealing a significant paradox between perceived learning and actual academic performance. Led by Somi Joo, Professor Changjun Lee, and Professor Daeho Lee from the Department of Artificial Intelligence Convergence, the investigation involved 88 university students who completed tasks equivalent to domestic high school academic achievement levels. The experiment was structured around three learning phases: concept understanding, problem solving, and result review, with participants randomly assigned to interact with GPT-4o at various stages. The data demonstrated that students utilizing generative AI during the problem-solving phase achieved higher scores on subsequent assessments. The artificial intelligence streamlined complex queries, reduced cognitive load, and eliminated redundant processing steps, thereby substantially boosting learning efficiency. However, the study identified a critical downside tied to overreliance. Students who frequently depended on the technology tended to perform worse on objective tests. This decline occurred because heavy users accepted AI-generated responses uncritically, bypassing the necessary cognitive engagement required to internalize knowledge. Interestingly, this drop in measurable achievement contrasted sharply with self-reported metrics. Highly dependent students rated their perceived learning and satisfaction significantly higher, attributing this to the AI’s ability to structure information clearly and provide immediate, digestible answers. Researchers describe this phenomenon as a learning illusion, where the perceived mastery of material does not align with actual comprehension. Professor Changjun Lee emphasized that while generative AI serves as a potent efficiency tool, uncritical dependency undermines genuine academic development. He stressed the urgent need for structured educational guidelines that cultivate independent questioning and critical verification skills. Rather than implementing blanket restrictions or promoting unconditional adoption, the researchers advocate for pedagogical frameworks that position generative AI as a supplementary aid. The findings provide a concrete pathway for integrating artificial intelligence into educational ecosystems responsibly, balancing technological leverage with cognitive rigor.