Peter Halley, best known for his geometric “conduit” paintings and neon color palettes, recently shared his first experiments with artificial intelligence. In an Instagram post, the artist wrote, “I asked AI to make a Peter Halley painting, but I think it needs more practice.” The images, of course, look more like a mess of bright colors and shapes than anything close to Halley’s distinct style. Yet the longer one stares, quintessential Halley fragments reveal themselves: his signature stripes in one, specific shades of pink and yellow in another. The images seem to portray Halley’s art as AI understands it rather than as Halley would. Hence the artist’s comment: “I think it needs more practice.” With AI (what I call machine learning via neural network architecture), the question becomes, is practice all it needs? In other words, given sufficient training data, could AI produce a Peter Halley painting (or something indistinguishable from one)? Is AI inherently creative the way humans are?
It is tempting to argue no, that AI is not and never will achieve true creativity (whatever that means). Rather, what it produces are mere algorithmic regurgitations of data that humans provide. Unlike humans who create based on a complex interaction between emotions and imagination, AI is limited to a kind of statistical repackaging. In this view, AI can produce novel poems and images, but because it doesn’t create the way humans do, it never understands these tasks as creative ones. While tempting, and not without some truth, this view more often than not flattens the significance of AI and how unprecedented of a technology it truly is. It should not be taken lightly, for instance, that Geoffrey Hinton, known as the “Godfather of AI” and winner of the 2024 Nobel Prize in Physics, frequently argues that AI is inherently creative and that it will create artworks and poetry better than humans. Hinton published a version of the backpropagation algorithm central to modern AI in 1986 and spent ten years working at Google leading AI research, eventually quitting so that he could “freely speak” about the dangers of his creation.
These claims put new pressure on the precise definition of creativity. Hinton’s definition is clarified in an interview conducted by the Royal Institute of Great Britain. He asks the audience to imagine asking an AI tool like ChatGPT how a compost pile is similar to an atom bomb. Hinton argues that most people can’t answer this; they haven’t thought about it or they think the two are very different. However, AI recognizes that although the time and energy scales are very different, what’s similar is that when a compost pile is heated, it begins to generate heat faster, and when an atom bomb produces neutrons, it begins to produce neutrons faster. According to Hinton, “AI has understood they’re both forms of a chain reaction and it’s using that understanding to compress all that information into its weights.” If it recognizes processes of change, it must also recognize hundreds of other analogies it sees but we don’t. This is precisely what creativity is: the production of “analogies between apparently very different things.” AI is not a mere regurgitation; it understands the same way humans do.
Recent developments in phenomenology and cognitive science seem to further solidify the link between AI and human processing. Anil Seth, cofounder of the Sussex Centre for Consciousness Science, argues in his book Being You that all conscious experience can be reduced to a kind of creativity (Seth uses the term “controlled hallucination”). According to Seth, as the brain receives sensory signals from the world, it continuously makes predictions about the causes of these signals. These predictions inform perceptual experiences at the same time that new signals reveal errors in past predictions. For example, the brain might receive signals that have in the past correlated with the specific experience of a cup of coffee and decide it sees a cup of coffee. Later, however, upon the reception of new signals, the brain might decide this prediction was incorrect; now the signals it receives more likely correspond to the existence of a cup of tea, so it decides it sees a cup of tea. The brain minimizes the errors of its predictions constantly through what Seth calls “prediction error minimization,” and this constant process forms the basis for the entirety of conscious experience itself. We never experience sensory signals themselves, only interpretations or creations of them. This is true not just for relatively simple acts like perceiving a cup of tea but also for more complex ones like feelings of jealousy or the internal creation of a sense of self. All conscious activity involves the interpretation of sensory signals via prediction error minimization.
Most AI follows basically the same process. Data consisting of both inputs and outputs is first “forward propagated” through a neural network. Weights and biases are assigned to “neurons” in that network such that its inputs are likely to achieve their target outputs. Then the network compares its outputs to the outputs in the original data set and assigns a “loss value” to each, quantifying the strength of the discrepancy. Lastly, the network “backpropagates” this “cost function” through the network again, allowing each neuron to adjust its weights and biases such that its errors are minimized. And this process—from data to forward propagation to backpropagation—forms the basis for all artificially generated outputs, whether that involves transcribing speech or writing a poem. AI re-applies the associations it “discovered” in its original data set to new inputs in the generation of novel outputs. It sounds a lot like the prediction error minimization we think humans do.
Still, the radically different material basis between something like ChatGPT and the human body cannot be overstated. There is a reason we call it artificial intelligence; its artificiality lies not in its intelligence, but in its materiality. Human creativity is always an embedded process, reliant on a human body and experience to exist, and thus always distinct from AI creativity. Importantly, this doesn’t mean AI is not creative or that it cannot produce something like a Peter Halley painting, just that AI creativity is at least materially different than human creativity.
Debates around AI are quickly muddled when we forget this distinction: between what something does and what something is. There is no reason to believe intelligence presupposes consciousness. AI is and will always remain a medium—a piece of technology—not a new life form. It is an extension of human creativity rather than its replacement. It was media theorist and philosopher Marshall McLuhan who first proposed this concept in the 1960s, of viewing media as extensions of bodily functions. Prior to McLuhan’s work, media were typically thought to be carriers of meaning. The photograph, film, or building was analyzed only to the extent that it channeled an intention on the part of an author. And while media do perform this function, McLuhan introduced another, perhaps deeper, impact: media’s effect on us. Media function as extensions of the body—the house as an extension of the ability to trap heat, or the photograph as an extension of sight. Media change how we relate to our environment, even as we change them. McLuhan famously argued that “the medium is the message,” giving media newfound agency in critical theory. The introduction of the camera not only allowed us to give permanence to what we see, it allowed us to actually see differently, in a new way. Technology is not a passive tool for human thought and action; it changes us as much as we change it. The invention of the house causes new structures in human relationships. The invention of the camera causes a new understanding of the frame.
AI is not exempt from this theorization. It is an extension of human creativity as much as it is a channel for human agendas. We are not immune from its inevitable reflexive effect on us. This is perhaps what’s truly revolutionary about it. Halley, after his original post, shared a follow-up in which he asks ChatGPT why making a Peter Halley painting has nothing to do with artificial intelligence. The question, in presupposing its answer, betrays Halley’s anxiety about the artist’s role in creation. Halley speaks about AI but in many ways echoes the general anxiety that has historically emerged with the invention of other media. Photography in the early 19th century forced artists to question their longstanding role as society’s chief means of image creation in radical ways. Artists asked that if a machine, the camera, reproduces reality more perfectly than the human hand can ever hope to, where does that leave them? It is not a coincidence, then, that abstraction in visual art emerged around this time. To reassert their value, artists had to shift. No longer just a means to replicate reality, painting turns inward, investigating its unique material properties and forms of expression.
The difference with AI, however, is that it is no longer the reproduction of reality under attack, but the entire process of creation itself. If machines gain the ability to create poems, images, and essays better than humans, where does that leave us? It might very well be that making a Peter Halley painting is different than asking artificial intelligence to, but this doesn’t mean AI won’t soon create something indistinguishable—or better than one. Just as an artist minimizes the errors of their sensory predictions in deciding which brushstroke to lay next, AI will minimize its cost function in deciding which bit to fill next. It is not hard to imagine an advanced machine that recognizes the processes of change underlying Halley’s depictions and relies on these to produce new iterations. As training data gains this level of sophistication, however, I suspect humans will create entirely inconceivable alternatives. The human body remains unique, although, in the end, Halley might be shocked to discover that painting and artificial intelligence have much more to do with each other than he cares to admit.
Image: Marsden Hartley, Provincetown, oil on composition board, 1916, (Art Institute of Chicago, Chicago, Illinois).