AI models developed by a KAIST research team performed improvised jazz with human musicians at South Korea’s Jarasum Jazz Festival. The striking part is that they responded to the music as it happened, rather than simply playing a prepared track.
Watch Desk analysis
What happened
At the festival, held from 9 to 11 October in Gapyeong County, 12 human musicians collaborated with AI across 26 pieces in 12 performances, according to The Korea Times. The team, led by KAIST professor Nam Ju-han, demonstrated two systems: Scott, an AI pianist, and DeepPocket, an accompanist.
Scott compared live performances with sheet music, then generated solos based on the musicians’ melodies and rhythms. Its music was played on a Yamaha Disklavier, an acoustic piano that reproduces the notes through real keys and pedals. DeepPocket listened to performers and generated piano, bass and drum accompaniment, adjusting to their notes, rhythm and tempo without prior preparation. The show also featured an augmented-reality pianist overlaid on live camera footage.
Why it matters
Improvisation is a tougher test than asking AI to play a fixed arrangement. Musicians change tempo and melody in the moment, so a useful partner has to listen and respond. The demonstration puts that interaction on stage, with people rather than a neatly edited studio track providing the musical cues.
The team says its aim is to develop AI as a musical partner, not merely a performance tool. That is an ambition, not a verdict on whether musicians will welcome the arrangement. But the practical demonstration gives the debate something better than another abstract promise: a system responding to live players across a festival set.
Our read
The interesting question is not whether a machine can produce notes. It is whether musicians can make something worthwhile with a system that listens, adapts and leaves room for the performance to change. Scott and DeepPocket offer a vivid test of that idea. The next measure should be what musicians make of the experience, not simply how many pieces the AI played.
What to watch
- Whether musicians describe the systems as useful collaborators in live performance.
- How Scott and DeepPocket handle changes in tempo, phrasing and improvisation across different groups.
- Whether the team shares further detail about how the models listen and generate music in real time.
Discussion spark: In live music, should AI be judged by whether it helps musicians create something new, or by whether audiences can tell it is there?
Sources and evidence
- Source update (16 October 2026, 02:05 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.