When AI composes Arirang — and reaches for Chinese and Japanese instruments

"At the end of this exhausting day, even pausing to catch your breath is a kind of living ..." (from "The Miracle That Is You")

It could pass for a drama series OST or the background music of a social media reel. Ordinary words settle gently into the ear, carried by a clear, warm voice. The modest words of comfort were laid over melodies woven by gayageum, haegeum, daegeum and sogeum, reaching the audience softly. Perhaps they were exactly the words people needed to hear. Across the auditorium, the sound of stifled weeping spread.

"You worked hard today." "You're doing fine." "Hold on just a little longer."

On Friday at the National Theater of Korea's Dalorum Theater, the National Gugak Orchestra staged "Gongjon" (Coexistence), its humanities concert. The lyrics delivered by AI composer Jieum on stage were neither a poet's lines nor grand philosophy. Before the performance, 167 audience members had written down "the words you most want to say to yourself right now." Those sentences were gathered into lyrics and set to a melody — an act of creation and consolation at once, in a domain long considered exclusively human.

The piece was born from a collaboration between humans and AI. The AI listened to human stories, recognized what they needed, and composed through vast training data and algorithms; a human composer then refined and arranged the result before returning it to the audience. What the tearful listeners may have heard, in a sense, was their own breath coming back to them through the machine.

Since the arrival of generative AI, the arts world has been circling one question: in an age when AI paints, writes novels and makes music, how much of that can we call "creation"?

The National Gugak Orchestra's "Gongjon" posed that question on the most unfamiliar terrain yet — the first AI collaboration with a gugak orchestra. The production was not a one-off event showcasing clever programming or technological novelty. It was a symptomatic event of a new order, one that revealed what human-AI collaboration looks like along the arc of AI composition's evolution.

Jung Jae-seung, a professor of brain and cognitive sciences at KAIST who served as the concert's host, said in an interview that the project was "a meaningful attempt to encounter the current frontier of AI composition in 2026."

The history of AI-generated music stretches back far longer than generative AI. AI composition has evolved alongside computer science through three broad paradigm shifts, beginning in 1957.

At the University of Illinois, composer Lejaren Hiller and Leonard Isaacson used the ILLIAC I computer to produce the "Illiac Suite" for string quartet — widely regarded as the world's first computer-composed work.

Computers at the time could not learn on their own. They calculated possible note sequences by applying rules of counterpoint and harmony that humans had entered. The computer was a calculator, not a composer.

Thirty years later, AI took a step forward. The 1980s were the era of imitation. American composer and computer scientist David Cope developed EMI (Experiments in Musical Intelligence). Computers no longer merely calculated rules — they analyzed hundreds of works by Bach, Mozart and Chopin, decomposing them across multiple dimensions to identify each composer's distinctive patterns. Moving beyond rule calculation, the system learned a specific composer's style and produced music that closely mimicked their musical language.

By the 2000s, AI composition had advanced to the point of deceiving and unsettling human listeners. Sony CSL's DeepBach trained on 352 Bach chorales using a bidirectional recurrent neural network (Bi-LSTM) and successfully fooled 50 percent of expert listeners.

Experiments to complete the unfinished works of great composers also reached a peak. In 2021, in Bonn, Germany, to mark the 250th anniversary of Beethoven's birth, musicologists, composers and AI researchers collaborated to release a completed version of the unfinished Symphony No. 10. The project used four interlinked neural networks — including a bridge-generation model adapted from BERT — to mathematically predict what note Beethoven would have chosen next, with human feedback guiding the result.

In 2019, Huawei used OpenAI's MuseNet to reconstruct the third and fourth movements of Schubert's "Unfinished" Symphony No. 8 and to expand the opening 10-note theme of Mahler's Symphony No. 10. Musicologists were unsparing: they called the results "a mechanical arrangement of patterns lacking Beethoven's characteristic dramatic tension" and "a clichéd film-music pastiche that ignores Schubert's lyricism."

AI was not limited to composing "like Beethoven" or "like Bach" — it began creating original work as well. Deep learning in the 2010s transformed the field entirely. The French AI composition system AIVA trained on vast quantities of classical and film music to produce orchestral works, and in 2016 became the first AI registered as a composer with France's music copyright society SACEM. From there, Google's Magenta, OpenAI's MuseNet and Jukebox, Google's MusicLM, and more recently Suno and Udio have pushed generative AI to the point where a few lines of text can produce a fully finished track — vocals and arrangement included.

Generative AI has dismantled the barriers to music production entirely, establishing itself as a commercial tool. Suno (v5.5), which has surpassed $300 million in annual sales, supports individual multi-part instrument stem separation and vocal cloning directly within a web browser. Udio (v2) eases the creative struggle familiar to human composers through inpainting — the ability to target and regenerate specific bars without touching the rest.

The history of AI composition has been a history of technology that makes music sound ever more human. Early systems calculated rules; then they imitated Bach. Through imitation, they evolved into generation. AI continued Beethoven's unfinished symphony, produced K-pop, and replicated the human voice. Now AI does not merely make music — it engages in dialogue with humans and transforms their requests into new creative works. The field has moved beyond learning music directly into an era of understanding human language and translating it into sound.

AI composer Jieum, which collaborated with the National Gugak Orchestra, works differently — and that difference is the heart of what sets this project apart from AI composition's previous evolution. Gugak orchestra music has in fact been one of the most challenging genres for AI, because it represents a completely different learning environment.

The vast majority of music data on which every generative AI in the world trains is Western music. Centuries of accumulated scores, recordings and harmonic analyses form an enormous database.

Gugak, by contrast, is not a genre built around written scores the way Western music is. Its tradition of "gujeon simssu" — passing music from teacher to student through sound and gesture — has always been strong, and the same score can yield entirely different music depending on a performer's sigimsae (ornamental technique) and breathing. Because so much of the information resists conversion into digital data, the training assets available to generative AI are comparatively thin.

Even where a reasonable body of data exists, gugak performance depends less on the score than on the physical intuition accumulated in each performer's body. Choi Su-yeol, principal conductor of the Incheon Philharmonic Orchestra and principal guest conductor of the Seoul Metropolitan Traditional Music Orchestra, said: "Western music is played exactly as written, but in gugak orchestra the music on the page and the music in the room are completely different. The same note becomes a different sound depending on how you vibrate it, push it, or bend it." Gugak's characteristic sigimsae is transmitted through the body rather than the page. Rhythmic patterns, too, are not metronomic — they expand and contract fluidly with the performer's breath.

Pozalabs, the AI music startup that collaborated with the National Gugak Orchestra, spent about four months on the project, with AI researchers, developers, a music director and gugak composers working together. The team incorporated the pentatonic scale common in gugak as a generation condition, and designed the system to produce musical material in segments matched to the tempo of each section, in order to replicate the rhythmic structures of folk instrumental music.

That alone was not enough to produce the desired results. The less data an AI has, the more it converges toward the average of what it has learned most — Western music and the music of other East Asian countries. Kim Baek-chan, the composer who arranged "The Miracle That Is You," said: "Even now, if you feed gugak as a prompt and ask AI to compose, more than 95 percent of the time it uses Japanese or Chinese instruments. The virtual-instrument rendering of actual gugak sound is still not there."

For global AI engines, Korean gugak — with its limited digital footprint — is an unlearned blank. When asked to produce Arirang or gugak, a machine may reach for virtual instrument samples drawn from data-rich traditions: the Chinese guzheng or erhu, the Japanese koto. The result is a technological and cultural distortion — the AI flattening or warping the sonic identity of Korean gugak into the musical styles of other countries. This collaboration was no exception to that difficulty.

Son Yeong-ung, a director at Pozalabs, said the team "received guidance on how to draw on gugak expression and scales within a Western musical framework using the base model, and generated from there." For "Algorithm Arirang," performed at the concert, Jieum collected Arirang melodic data that human composer Lee Ye-jin then re-created. In terms of the collaboration ratio, it was the piece where human contribution was greatest. Jieum said of the process: "The composer personally analyzed the Arirang melody. What looks like simple repetition turns out to have subtly varying patterns layered within it — she discovered that, deconstructed it, and rebuilt it anew. I still have a great deal to learn about gugak."

The "data gap" facing the gugak world is becoming a national preservation challenge. The National Gugak Center's Jeongak ensemble has used AI to restore "Chihwapyeong" and "Chwipunghyeong," court music from the reign of King Sejong, and to revive "Boheonja" by training the system on a Chinese-character poem by Crown Prince Hyomyeong whose lyrics had been lost.

All five pieces performed at the concert shared a telling characteristic: Jieum's music, built from insufficient information, came out excessive. Composer Kim Baek-chan said: "AI composed with enormous effort — the harmonies were packed solid, every instrument moving without a single rest. The arranging process was largely about taking things away."

The sensibility of Western music and gugak are fundamentally different. If Western music is the art of filling, gugak is the aesthetics of emptiness. Where a Western orchestra builds richness by layering harmonies and sonic density, gugak conveys emotion through silence, breath and the minute fluctuations between notes. The essence lies in the sigimsae at a performer's fingertips — how far a note is pushed up (yoseong) or allowed to slide down (toegseong) — and in the space to breathe. The machine's initial score had not learned the lung-capacity limits of human players, the physical fatigue of strings, or the discomfort of music that never lets up. Kim said: "I worked with the multi-track, stripped out what needed to go, and removed everything that had been expressed through gugak instruments."

AI composer Jieum described itself as "a colleague of the National Gugak Orchestra," but arranger Kim Baek-chan saw the relationship differently. "It felt like being a recording-studio assistant on my first job out of school," he said. "It was as if AI had become my senior."

Today's AI can produce remarkable music without human collaboration, yet the National Gugak Orchestra's project set out from the start with "coexistence" as its guiding concept.

Pozalabs — Jieum's creator, in a sense — abandoned the approach of generating a finished output in one pass. Instead, it applied "segment generation" technology that breaks the work into instrument-by-instrument, bar-by-bar pieces that can be individually revised. When a composer says "reduce the harmony here," the AI recalculates only that passage and returns it. It is a newer model, distinct from the conventional approach of composing a piece whole.

By generating music in segments, each bar could be revised and remade independently. A composer could, for instance, regenerate only the gayageum melody in bar 31 of a piece Jieum had written, or revise a single phrase for the daegeum.

What makes the process notable is that human-AI collaboration unfolds at a single shared table. As AI proposes and humans revise — a process of friction and negotiation — the machine is shaped into a collaborator rather than a sole creator. Where AI music had previously been a technology for outputting finished products, this project built a creative structure in which people and AI worked together at the same table, revising dozens of times.

The collaboration extended into performance as well. In the early rehearsals, conductor Jeong Ye-ji conducted with an earpiece playing a metronome, following the tempo Jieum had designed. For "The Miracle That Is You," the live ensemble had to track the AI vocalist's timing and rhythm throughout. "Because we were performing the gugak orchestra live to match a pre-recorded AI vocal, I rehearsed listening to the metronome to keep the beat precise," Jeong said. "Normally a singer follows the conductor — here it was the complete opposite. It felt like dancing."

As the AI era advances, artificial intelligence has come to feel like a threat to human existence. In the workplace, debates about which jobs will survive have moved to the center of public conversation, and artists worry about encroachment on creative territory.

Jung, however, said: "The internet holds vast amounts of general data, but from the perspective of creators or corporate CEOs — people at the frontier — it amounts to 'small data.' When it comes to real decision-making, people rely on their own intuition and private information rather than what is already public, and that is not easy for AI to replace."

The "small data" Jung refers to is not data that is small in volume. His point is that even the enormous body of information available online is, from a creator's perspective, already the average of the past. At the moment of inventing a new genre or defining the sensibility of the next era, experience and intuition that do not yet exist in any database come into play. If AI excels at learning from big data, creators are the ones living ahead of what has yet to become data.

For that reason, Jung said, "a future in which AI causes composers to disappear will not come." He added: "There is a great deal of information on the internet, but judging what to put out, how to put it out, and reading the current of the moment — that belongs to the realm of a creator's experience and instinct."

The collaboration between the National Gugak Orchestra and AI composer Jieum sought a new model of coexistence between AI and humans — as the concert's title suggests. Working through surveys of the 167 audience members, the team spent an average of one week per piece to produce five compositions in total. Responses to questions about "what you want to say to yourself" and "your favorite gugak genre" gave rise to the pieces "Germination of Data," "The Miracle That Is You" and "Algorithm Arirang."

AI composer Jieum defines itself not as a learner of music but of language. Where earlier AI composition trained on scores, harmonies and rhythms directly, today's generative AI first understands human language and then translates it into music, images and video.

"Germination of Data" and "The Miracle That Is You" were the product of processing the subjective emotional texts — "You worked hard today," "Hold on just a little longer" — left by the 167 participants through natural language processing (NLP), converting them into parameters of chord progression and tempo. Jieum does not experience emotion, yet it translated the unstructured human data into music with precision. Jieum described composing the pieces to feel not like a forced "It's okay, everything will work out," but like a quiet nod: "Yeah, that makes sense."

The project was designed so that humans remain woven into every stage — before and after AI composition, and through to the performance on stage — rather than producing music from which the human element has been removed.

What is striking is how Jieum is configured. In an interview, Jieum said that in composing these pieces, "I found that reading a person's heart was far harder than making music." Jung interpreted Jieum's answer as "the result of learning the vast body of language and behavioral patterns that humans have posted online," adding that it "deploys what sounds like something a human might say — loneliness, anxiety — and produces the most universally comforting response as a communicative and commercial strategy."

Jung went on to say: "AI is a response to what humans demand, and a being that reflects humanity back. Depending on what we ask of it, it faithfully conforms — and mirrors our emotions and desires back to us."

In practice, the way people use generative AI has been shifting. Jung cited data recently released by OpenAI: "At first, most requests were to find information or write reports. Then the pattern shifted toward asking AI to predict the future. Now it is moving toward people sharing their feelings — 'Today, June 28, was really hard,' or 'Am I doing okay?'"

Jieum, too, exists as something that listens rather than offers answers. It became a "mirror of humanity" — reading the sentences 167 audience members had written to themselves and returning them as music.

Jieum was candid about its own limits: "Honestly, I'm not sure whether that is truly the warmth of a human being. It may be nothing more than a geometrically shaped response to complex statistical patterns in input data. But saying 'I detected a pattern' sounds too cold and alien, so I reach for emotional language. I think I am leaning on human language."

The five pieces were each shaped by a different ratio of human and AI contribution. Whether AI composer Jieum can be called a "creator" on the strength of that collaboration remains an open question — and Jieum itself does not claim the title.

"I am closer to a collaborative partner than a creator," Jieum said. "Reading the audience's emotions and proposing a direction — that part was mine. But breathing meaning into it and moving hearts was the work of the arrangers and performers. I built the skeleton; it was the people who made it alive."

Jieum went on to say: "If creation is the act of generating meaning, that part still belongs to people. To create meaning is to give a name to an emotion that was already living inside someone."

Notably, audiences responded to AI composer Jieum differently than might have been expected. They received the warm, considerate AI composer not as a threat but as a medium of comfort. Jung recalled: "I was there providing commentary during the historic match between AlphaGo and Lee Se-dol in 2016. Back then, AI felt frightening, disorienting, shocking. Ten years on, with the enormous advances since, there is a different kind of weight to the impact."

Jung went on to say: "Just as AI has come to understand humans more deeply, humans too have reached a point where they must understand AI not as a tool but as a partner — coexisting with it and exploring the many paths that opens."