Abstract
Although artificial intelligence is fundamentally reshaping the ecology of music learning, existing research has disproportionately emphasized performance outcomes while underexamining psychological mechanisms, leaving the tension between technological empowerment and cognitive dependence theoretically underarticulated. Following PRISMA 2020, we systematically searched four databases and included 21 empirical studies to examine how three AI tool types-assessment-oriented AI, generative AI, and Comprehensive/adaptive AI-differentially shape learners' self-beliefs and cognitive agency in music education. The evidence base remains geographically and developmentally concentrated: most studies were conducted in China and in higher education, while early childhood settings were absent. Using thematic analysis, we conducted cross-type comparisons and synthesized psychological pathways. Assessment-oriented AI most consistently strengthened ability beliefs via objectified, visualized feedback and positioned cognitive agency around self-monitoring, self-reactiveness, and self-reflectiveness. Generative AI tended to enhance value-attitude beliefs and intentionality by lowering technical barriers and reconfiguring learners' creative roles toward aesthetic decision-making and output curation. Comprehensive/Adaptive AI more often supported forethought and sustained engagement by dynamically maintaining alignment between task challenge and learner capability. Across studies, psychological empowerment manifested as increased perceived competence and control, heightened motivation and engagement, and visible self-regulated learning behaviors. Cognitive dependence, however, emerged through outsourcing evaluative authority, score-driven goal distortion, algorithm-accommodating self-censorship, and attributional shifts that tether confidence to technological support. Developmental differences were also observed regarding dependence mechanisms: primary learners tended to perceive AI as a restrictive "scoring referee," whereas higher education students demonstrated strategic agency in orchestrating AI assistance. Specifically, a critical construct-tool mismatch was identified: while assessment AI consistently supports self-reflectiveness, generative AI currently lacks sufficient evidence for fostering learners' forethought. In light of the identified construct-tool mismatch, future research should prioritize addressing the paucity of evidence on how generative and adaptive AI foster forethought and intentionality, thereby clarifying whether such technologies ultimately reconstruct or erode learners' cognitive agency.