Artificial Creativity
This conversation is part of the AI & Creativity podcast series, where I talk with the researchers, artists, and founders shaping how machines and human creativity meet. To open the series, I sat down with Ali Nikrang — key researcher and artist at the Ars Electronica Futurelab in Linz, and a professor at the University for Music and Performing Arts Munich, where he works directly with composition students on integrating AI into their craft. Nikrang has done something most people still assume is impossible: he’s used AI to compose orchestral music performed live by ensembles like the Munich Philharmonic. But his real argument is more provocative than “AI can write symphonies.” It’s that the tools everyone is excited about — the prompt-to-song generators — are the wrong tools entirely for serious artists.
From Bach’s Canons to Mozart’s Dice: Algorithmic Music Is Centuries Old
One of the first things Nikrang does is dismantle the idea that algorithmic composition is new. Composers have leaned on formal systems for centuries.
He points to J.S. Bach’s contrapuntal works, where a single theme is played against itself in reverse (retrograde) or turned upside down (inversion) — rigid mathematical structure that nonetheless produces deep emotional beauty. Then there’s Mozart’s Musikalisches Würfelspiel, or “dice game”: Mozart wrote 176 pre-composed musical fragments, and performers rolled dice to determine the sequence, generating thousands of unique, grammatically correct waltzes centuries before digital computing.
Nikrang’s own path into this work started in 2005, reading Douglas Hofstadter’s Gödel, Escher, Bach. It planted a question he couldn’t shake: why couldn’t powerful computers write music that sounds genuinely human? At the time, rule-based systems couldn’t capture the implicit nuances of music — a limitation only modern probabilistic AI models would eventually overcome.
Can AI Compose Classical Music?
Yes — modern AI can generate complex, natural-sounding classical music, and Nikrang proved it in 2019 by using machine learning to compose completions for Gustav Mahler’s unfinished 10th Symphony. The point of that project wasn’t to “finish” Mahler. It was to demonstrate a threshold moment: that a model could produce convincing late-19th-century orchestration at all.
The engine behind much of his work is Ricercar, an AI composition platform he began building at the Ars Electronica Futurelab in 2019, developed in collaboration with music universities. The name is deliberate. A ricercar was a Renaissance and Baroque musical form in which composers methodically explored the contrapuntal potential of a theme — and the Italian word itself means “to search out,” the root of the word research. That etymology is the whole philosophy: Ricercar is not a fast-composition button. It’s an interactive tool for exploration, letting composers manipulate the rules the AI learned in order to discover unconventional musical paths.
Why Text Prompts Fail for Serious Music
Here’s Nikrang’s sharpest and most counterintuitive argument, and it cuts against the entire current wave of prompt-to-music apps: text prompting is fundamentally broken for complex music generation. He gives three reasons.
First, a vocabulary gap. Image generation thrives on text because visual reality has tens of thousands of concrete nouns and adjectives — “glasses,” “foggy,” “a dark wooden table.” Music has almost no precise verbal equivalents for structural, harmonic, or textural ideas. As he puts it, “how can you communicate verbally about music? It’s a conceptual problem and a philosophical problem.”
Second, emotional subjectivity. Prompt words like “happy” or “melancholic” collapse because listeners categorise the same piece into wildly different emotional buckets. There’s no reliable mapping from the word to the sound.
Third, and most fundamental, the constraint of time. A text prompt slaps a single static label onto an entire piece — but music unfolds second by second. A 30-minute movement is constantly changing, and no fixed prompt can steer that evolution.
His alternative is what he calls musical inspirations: instead of typing words, a composer feeds the AI a reference piece or snippet, and the model analyses its texture, harmonic progression, and movement to guide what it generates next over time. Guidance becomes musical and time-based, not verbal.
The Real Goal: Breaking the Model’s Predictability
If an AI is trained to predict the most probable next note, it will, by design, produce conventional, predictable results. For an artist, Nikrang argues, that’s useless. “I press the compose button and I get something that sounds very conventional,” he says — and in that moment, “what’s my role actually here in this process?” A predictable output leaves no room for human identity.
So the genuine technical and artistic challenge is the opposite of automation: it’s forcing the model out of its comfort zone. The aim is to extract the structural understanding the model absorbed during training and recombine those learned features in non-standard ways — producing music that still respects musical logic without collapsing into cliché. Fascinatingly, his composition students sometimes hit the reverse problem: the AI is too unpredictable, and they have to “negotiate” with it to rein its scope back in.
Is AI-Assisted Music “Real” Art?
Nikrang’s answer is unambiguous, and it’s the thread that ties this episode to the larger question of AI abundance: co-created work is fully human work. “If you are collaborating with AI, if you are co-producing something with AI, at the end of the day, it’s human works.”
His reasoning is historical. Artists have always used mechanics, complex tools, and abstract systems to make art — Mozart’s dice included. When a musician works with AI, the intent, the curation, and the guidance stay entirely human. The anxiety about “fake art,” he suggests, dissolves once the tool enables genuine personal expression, because the focus returns to the artist’s choices rather than the technical backend.
What machines still lack is the one thing that can’t be engineered: intention. “The AI systems we have right now are all very good, of course, but they don’t really have the intention to create something new,” he says. “Intention would be the answer.” A model has no intrinsic desire or purpose to create. That gap is precisely where the human remains irreplaceable.
What This Means for Musicians and Creators
Three practical principles come out of the conversation:
Steer with music, not words. Use structural, non-verbal inputs — MIDI textures, reference audio — to guide AI generation, rather than leaning on text prompts that can’t capture what you mean.
Refuse the first output. Don’t accept the most predictable result the model hands you. Adjust the interaction until the AI recombines its knowledge in unconventional ways.
Treat AI as a mirror. Use its outputs to expose your own habits and clichés. An AI that echoes your default moves back at you is a tool for understanding — and escaping — your compositional reflexes.
Watch the Full Conversation
Nikrang’s perspective is a useful corrective to the hype: the future of AI in music isn’t a machine that composes for you at the press of a button. It’s a collaborator that pushes you somewhere you wouldn’t have gone alone. As he puts it, artists are “anyway forced to explore new ideas — we can’t do the same as the generation before us.”
The full interview goes deeper into Ricercar, the Mahler project, and why intention is the line machines can’t cross. Watch it on the LiveInnovation YouTube channel.
