What happens when artists break AI?
Artificial intelligence is neither an artistic movement nor the end of art. It is a tool, although tools have a habit of arriving with claims of cultural catastrophe attached. The sampler once appeared to offer musicians something equally alarming: a machine capable of capturing any sound and reproducing it on command. Why employ an orchestra, record a drummer or learn an instrument when reality itself could be stored in a digital bank? The objections sound familiar. This was not music. It was theft. The machine was doing the work. Yet the sampler became interesting when artists stopped treating it as a cheaper imitation of what already existed. Art of Noise exploited the Fairlight’s limitations, throwing fragments of voices, machinery and orchestras into forms their original owners could never have anticipated. Its imperfections were part of the attraction. The technology did not simply reproduce reality. It damaged, rearranged and released it.
The same history extends much further back. Jimi Hendrix reversed tapes, making the recording studio behave in ways its designers had not intended. Photographers turned a mechanical process into an art after painting had spent centuries pursuing increasingly persuasive illusions of reality. Even that division is less stable than it appears. Long before photography, artists used devices such as the camera obscura and camera lucida to project or reflect the visible world for tracing. Technology did not suddenly intrude upon a previously pure human activity. Art has always involved instruments, systems, borrowed knowledge and shortcuts. The arrival of a new machine merely makes those dependencies visible again. There are serious questions about consent, copyright, employment and the material used to train generative models, but declaring that AI cannot produce art does not answer them. A tool can be used unethically without every possible use of it becoming artistically invalid.
Most AI music currently makes the weakest possible case for its own existence. It is asked to generate a song that resembles a song, then praised for passing as something a human might have made. This is technically impressive and culturally tedious. A figurative painter and a child pointing a camera at the same subject have not performed the same act merely because both produce an image. Intention, selection, context and transformation still matter. Typing a conventional request into a model and accepting its first response may be authorship in the most minimal sense, but the interesting question is not whether it qualifies. It is what the artist has discovered that could not have emerged through more familiar means. Perfect imitation is usually where a medium begins, not where it finds its purpose.
Some artists were already moving beyond imitation before generative AI became a consumer product.
Arca used the Bronze system to create 100 evolving versions of “Riquiquí”, replacing a definitive recording with music able to rearrange itself. The same technology powered Echo at MoMA, where her composition changed continuously in response to activity taking place inside and around the museum.
On “Godmother”, Holly Herndon trained Spawn, a neural network built from human voices, to respond to producer Jlin’s music. Her later project with Mat Dryhurst, The Call, transformed the creation, performance and collective ownership of an AI training dataset into the artwork itself.
In “God Is An Algorithm”, Alfredo Violante Widmer used early AI models to generate melodic sequences with unexpected modes alongside abstract machine-generated imagery. He treated their instability as creative material, allowing the model’s strange decisions to redirect the human composition instead of completing it.
For Chain Tripping, YACHT trained machine-learning systems on their own catalogue, then selected, rearranged and performed the unfamiliar material returned to them. Songs including “Loud Light” used AI to disrupt the band’s established habits instead of producing a finished imitation of their style.
The real pioneers of AI music may therefore be absent from Spotify’s endless feed of synthetic songs. They are more likely to appear at Tate Modern, the Serpentine, MoMA or Berlin’s experimental art institutions, where music can become an installation, an argument or a system that refuses to produce the same result twice. Perhaps an artist will reverse AI as Hendrix reversed tape, feeding its answers back into itself until recognition collapses. Perhaps prompts will be written as poetry, sabotage or deliberate misunderstanding. Perhaps the traditional relationship will be inverted entirely: a human musician will compose for an artificial audience whose reactions alter the work, leaving us to watch a machine decide what it believes it has heard. We cannot predict the decisive misuse because, by definition, nobody has discovered it yet. But that is where AI will become culturally interesting. Not when it finally learns to sound exactly like us, but when an artist makes it do something none of us thought to ask for.