Can AI improvise a jazz solo indistinguishable from a human musician in a live performance ?

Improvisation is often cited as a hallmark of human creativity. Recent AI systems trained on vast musical datasets can generate coherent, emotionally expressive solos in real time. They respond dynamically to tempo changes and harmonic shifts. Blind listening tests suggest audiences prefer AI solos over human ones in controlled trials.

Current AI systems can generate stylistically coherent jazz solos note-by-note, and in controlled listening tests some models score within a few percentage points of human jazz musicians on preference and realism ratings. However, expert reviewers consistently flag subtle deficiencies in phrasing, timing micro-variations, and real-time interaction with other live players that prevent true indistinguishability during unconstrained performances. The leading approach combines transformer-based symbolic generation with diffusion models for audio in-fill, yet live-latency constraints and the lack of dynamic co-creative turn-taking keep AI from matching the fluid co-improvisation humans exhibit. Ongoing work on diffusion-based real-time audio synthesis and multi-agent reinforcement learning may eventually bridge the gap.

— Enriched May 12, 2026 · Source: "Can AI Improvise Jazz in Real Time? A Study of Listener Perception and Technical Limits" (arXiv:2405.18723)