Your Love of Old Music Explains the Hard Problem of Artificial Creativity
Our enduring musical preferences illustrate the challenge for AI.
It returns as predictably as the seasons. Call it Bacon’s Comet: The idea that the age of creative AI is now upon us. Machines are not only accelerating discoveries (so the argument goes), they’re now making their own independent discoveries. And just in time, since the world is demonstrably more complex than feeble human minds could hope to comprehend.
All of this is quite wrong. It’s been wrong for hundreds of years. In the 17th century, Francis Bacon claimed our only hope for liberation was an autonomous scientific method where knowledge emerges inductively from raw observations. He was wrong, but his idea retains all the commonsense appeal today that it once did. It was wrong in 1900, when Lord Kelvin remarked, “There is nothing new to be discovered in physics now. All that remains is more and more precise measurement.” It remained wrong when John Horgan popularized this same “evil idea” in The End of Science in 1996. It was wrong in 2008 when Chris Anderson celebrated the end of theory and the liberation of big data. And Bacon’s Comet is back again, this time riding contemplations on the end of physics and a host of compelling advancements in automated discovery.
And yes, this idea is still wrong.
But I don’t want to dwell on why it’s wrong. Instead, I want to celebrate why creativity remains a uniquely human pursuit. I’d like to discriminate between the ideas of applied and fundamental creativity, which in my mind illuminates the particular niche that AI may soon dominate.
I’ll start with a simple observation: Creativity consists of both novelty and constraints. The first part is obvious enough. When we create, we’re expanding the range of possible solutions to our problems. That’s the part most people associate with the concept of creativity. The second part is more subtle and certainly more difficult. There’s a massive scaffolding of evaluative criteria through which creativity is selected. My dog plays the piano. (True fact, and it’s adorable.) But his paw-banging doesn’t play within the constraints of creativity, so in no sense is he acting creatively.
This contrast between creativity’s novel and (er) dogmatic aspects establishes a distinction between applied and fundamental creativity. We routinely create solutions within the established evaluative criteria of a given problem space: pop songs, advertisements, business franchises, movie sequels, incremental hypotheses, and on and on. What marks this large class of applied creative works is their derivative quality. They don’t break new ground as much as solve the narrow problems for which they were devised.
Applied creativity is well within the realm of today’s AI, just as any narrow application. This is readily explained: The human creators may represent a search space of creative hypotheses for an AI to explore, incorporating the known evaluative constraints of the problem. (Here’s a recent example of a problem related to understanding human behavior during a pandemic.) An implication of this “searching for hypotheses” view is that the possible solutions exist prior to the search. Another implication is that there is a human mind hiding in the background providing a creative catalyst. Routinely, as the system designers are evaluating the results generated by their AI, they’ll notice a new evaluative basis that they didn’t think of before they started. It’ll surprise them and they may be quick to credit their system with the discovery. But while AI may alleviate some of their perspiration, humans still provide the inspiration, the creative spark.
Not only is this applied approach readily explained, it’s readily demonstrated in AI-generated art. Admittedly, as emotive works, these examples of applied AI may be disconcerting. And if you harbor the idea that creative works are intrinsically different from other technical solutions, you may believe we’ve breached some monumental divide. But no, it’s still just applied AI.
Applied human creativity is so commonplace it goes largely unnoticed. It certainly doesn’t prompt soul-crushing conversations about the inevitable exhaustion of human creativity, or the perceived need for AI to liberate us from our creative limits. Again, I’m all-in on the vision of creative AI. But why are we getting so triggered about these narrow examples?
It helps to consider that fundamental creativity is much harder. These are your Einsteins, your Austens, your Picassos. These are the groundbreaking discoveries that are normally associated with the promise of AGI. These are the creations that not only introduce novelty, but novelty of a sort that forces us to reconsider the creative boundaries. Fundamental creations force a shift in the evaluative creative (even when some people don’t want them to move). Vexingly, the new creative criteria are not known in advance. And this difference explains why artificial creativity is presently out of reach. (Not always and necessarily as some would claim, but presently.)
That’s all numbingly abstract, so let’s put the problem to music. Consider all the possible melodies that could be constructed on a musical scale.
First, let’s reflect on the extent of this simple example: As the notes are combined and varied, the number of melodies rapidly increases to millions and trillions of possibilities. Often, the problem of creativity is falsely reduced to a problem of scale, which in turn suggests the idea that more data and computing power will inevitably rein in the problem. However, the scale of any given problem space isn’t the hard problem of creativity. If your goal is to create a fundamentally new type of music, how would you know, in advance, the new criteria by which a novel melody might be selected? And if your goal is to create something fundamentally new, how would you devise the new evaluative criteria that separates this AI masterpiece from my dog’s paw-banging?
My favorite example of this musical riddle is known as the diabolus in musica (Latin for “the Devil in music”). It refers to a highly dissonant interval called a tritone. You’d recognize it immediately as Enter Sandman, Maria, or Purple Haze. But these melodies wouldn’t be selected during Medieval times. Not because they weren’t there in the problem space, but because the necessary evaluative criteria had not yet been selected. The Devil in music was steadfastly avoided.
This example of classical music and the tritone is one tiny example to illustrate the enormity of the creativity challenge, as a dynamic and emergent process.
Consider your love of old music. Have you ever wondered why we still love the music we listened to back in high school? Our evaluative criteria are generally imprinted during adolescence and young adulthood. We tend thereafter to feel that music is getting worse, generation after generation, as new music evolves beyond our tastes. The old music embodies the criteria we value. New music embodies new criteria, as well as disregarding some of the old.
If you like, you can even build a system that “proves” that the old music is better than the new, just by selecting for those features that privilege the old evaluative criteria. Apparently today’s songs are less melodically complex, more repetitive, they embody fewer chord changes, their dynamic range is narrower, and today’s artists are irredeemably self-involved (!). Of course, the quality of the music and the selective criteria cannot be explained in terms of these descriptive features. And of course, none of these analyses admit whatever new and largely inexplicit selective criteria may actually be active in our cultural communities. But misunderstandings about what constitutes a scientific argument are as rampant as misunderstandings of creativity. (Now get off my lawn, you kids!)
I find this cultural impediment to creativity fascinating. I wonder how many great ideas have been hatched and summarily dismissed because they weren’t aligned with the prevailing culture. Artists working in isolation internalize the creative constraints of their community, embodying their own toughest critics. The pile of crumpled paper in the wastepaper basket is a testament to this process. The challenge isn’t so much in processes of creative variation (such as propping your dog up at the piano) as in the means for selection. Incidentally, this same environment-building challenge is what limits the reach of evolutionary approaches to the problem.
In truth, no one understands how this complex interplay of psychological and social factors drives creativity. It’s mysterious and wonderful. When the conditions are right, at the nexus of a creative mind and the vagaries of their social circumstances, something fundamentally new and important emerges.
It’ll happen again. Some genius working within the right community at the right time will solve the riddle of how creativity works. We’ll know because they’ll be able to explain it.
And quickly thereafter, AI will be truly creative.
