Artificial intelligence is rapidly permeating music education, from smart practice applications to compositional aids. Unlike other disciplines, however, music learning involves highly sensitive domains such as creativity, embodied practice, and personal expression, making the nature of human-AI interaction particularly delicate. Current AI designs often fall into a dichotomy: either becoming a cold "digital metronome," whose mechanical feedback can stifle student motivation, or a disingenuous "companion" whose hollow praise fails to support genuine growth. To resolve this dilemma, this paper proposes a new framework based on the concept of pedagogical scaffolding: the "Learning Alliance." Adapted from the "therapeutic alliance" in psychotherapy, this framework aims to establish a professional, trust-based partnership between AI and student, focused on mutual growth.The core thesis is that within the vulnerable context of music education, a healthy Learning Alliance must be built on a nuanced balance of cognitive and affective empathy. The AI should be led by cognitive empathy, precisely diagnosing a student's technical or expressive hurdles. Simultaneously, it must judiciously apply affective empathy to validate the student's effort and creative courage. This design philosophy does not seek to replace human teachers but to create a new form of pedagogical scaffold—one that supports and protects the deeply personal, exploratory, and emotionally invested journey of becoming a musician.