Tech Policy at Duke University

Before Impressionism, there was panic

In 1839, celebrated French painter Paul Delaroche dramatically announced: “From today, painting is dead.” Delaroche had just encountered a daguerreotype, the first publicly shared photographic process. It would be another 20 years before Claude Monet would paint his first landscape, yet painting was being pronounced dead in the face of a technology that threatened its replacement.

With the invention of the camera, the process of recording the world with precision could now be passed off to a new medium. It was around this time that visual art began to move away from realism and toward artistic movements centering perception and impression. The Impressionists, most notably Edgar Degas, were deeply influenced by the ways in which photographers captured sequential motion, and Degas’ technique of cropping grew out of his own practice of photography. The camera would continue to influence the experimentation of modernist artists, namely Picasso, whose photography practice inspired his melding of multiple viewpoints within the same painting. Best observed in Picasso’s “Houses on the Hill,” this technique became a foundation of Cubism. Ironically, the camera’s gift to artists was the threat it initially posed.

While there are observable parallels between this historical moment and the present-day threat posed to artists by artificial intelligence (AI), this analogy has limitations. The camera can reproduce the world as it is, but AI systems can simulate human creative expression in almost any medium through which we interpret it. The case that AI will prove an equal gift to the art community is far from obvious, especially with a threat to artists’ livelihoods that far outweighs the job displacement experienced by portraitists in the 19th century.

When the artist builds the system

For musicians, the term “generative music” existed long before AI systems could generate compositions capable of resembling the creative output of humans. Musician Brian Eno, a pioneer of the ambient music and electronica genres, explains that something as simple as a collection of wind chimes can be thought of as generative. Each chime is tuned to a specific pitch, and the arrangement of chimes is creatively constructed. However, the way the wind ultimately strikes the chimes to create music is beyond the designer’s control.

Eno explains, “You can’t really say that you composed that particular performance, though you can say that you built the system from which that performance emanated.” An artist can build a dynamic system that essentially places the wind in a position of authorship, which can result in the creation of sound that the artist could never have predicted. While human authorship is one step removed, the artistic process is abstracted, not replaced.

The creation of systems that breathe the unexpected into music is also a hallmark of computer music. John Supko, a composer and professor at Duke University, uses generative computer systems as collaborators in his compositional practice to disrupt and challenge his musical sensibilities. The computer algorithms he designs help him discover fortuitous musical sounds and ideas, but his pieces ultimately reflect his own creativity.

In 2014, Supko and his collaborator Bill Seaman released the album “s_traits,” which they describe as the “product of three minds: two human, one artificial.” Importantly, the artificial mind was not trained on the creative output of other artists, but was instead fed nearly 110 hours of compiled audio that Supko and Seaman recorded or composed themselves. From this library, the composers used software developed by Supko to generate multi-track compositions, sifting through hundreds of drafts to land on 13 sketches each that they would further shape into the 26 tracks for their album. The process of building the “artificial mind” required years of intentional design and creative reasoning, and arguably demanded the same intentionality as a traditional compositional practice. Stephen Ni-Hahn, Duke graduate and designer of interpretable music generation models, names the creative process of designing systems “meta-composing.” Ni-Hahn’s research explores how AI can be made transparent enough for artists to shape its output meaningfully. However, the human-centered approach to generative music modeled by Supko and Ni-Hahn remains rare, especially as text-to-music systems like Suno AI that are trained on the work of artists dominate the market.

What Suno AI can’t do

While Supko’s compositional tools output the unexpected, text-to-music AI tools, like Suno AI, are confined to producing the familiar. Degas didn’t use photography to produce more of the same paintings faster; he used the novel technology of his time to see the world differently. Beyond a small community of artists building generative systems from their own creative material, AI is being used to mass-produce generic music rather than transform how we listen. Type a prompt into Suno, for instance, “melancholic piano ballad, female vocals, radio-ready,” and it will return something that sounds like a cleanly produced amalgamation of what you would hear on the radio when Adele topped the charts, but with uninspiring lyrics and predictable harmonic moves.

Yet, the deeper problem with Suno AI’s statistical aggregation of human creativity is that Suno is trained using the creative material of composers and songwriters who never consented to this use of their work. The camera threatened the livelihoods of portraitists, and while it could observe the world as a human would, it never extracted labor from it.

From today, painting lives

In 1874, fifteen years after Delaroche declared painting dead, Claude Monet exhibited “Impression, Sunrise,” the painting that would give Impressionism its name. The movement was born of artists willing to ask what painting could do that photography never could. The musicians who stand to gain most from generative AI are asking a similar question. Supko, for instance, emphasizes that technology adds, rather than removes, challenge and complexity to his practice. The question isn’t how to use Suno to produce mundane music faster, but how to build systems that enable artists to hear what they might not have discovered.

That said, with music streaming platforms like Spotify increasingly turning to AI-generated music to undercut artists’ revenue streams, it would be remiss to frame the potential for another artistic renaissance as entirely in the hands of forward-thinking artists. In the 19th century, the art world was being pulled in two directions: between the artists who lamented what photography threatened, and those who used it as a lens through which to see the world differently. All Cubism needed was a creative thinker like Picasso, and an audience ready for something new. Today, music creation suffers from no shortage of creative ambition, and is instead being directed by two very different forces: those who wish to extract value from working artists, and those who wish to protect such artists. A working musician today who spends years building a generative system from their own creative material is doing so in a market that rewards volume rather than creative autonomy, and offers artists no seat at the table to negotiate their terms. The Protect Working Musician’s Act can change that by allowing artists to collectively bargain with the companies whose business models are built on their labor. AI has arrived in the music world as a Trojan horse, but only because of legal and commercial conditions that are unfavorable to artists. With policy intervention, AI can follow the historical pattern of art and technology, eventually becoming to musicians what the camera was for painters.

What you can do

The working musicians who will do the most interesting things with technology are already out there. They just need a fair market in which to survive. Support the Protect Working Musicians Act by signing this petition to help build the conditions for new and existing art forms to grow.