Do androids dream of electric beats? Photo: Prospect composite

Can machines create?

Artists and musicians beware—AI has got much further than you might think
March 4, 2019

In 2012, Iamus released a CD of classical music performed by the London Symphony Orchestra. Music critic Tom Service was somewhat dismissive, calling Iamus’s composition Hello World! “so unmemorable, and the way it’s elaborated so workaday, that the piece leaves no distinctive impression.”

But it wasn’t a bad debut really—when you consider that Iamus is not a composer but a computer algorithm, developed by researchers at the University of Málaga in Spain.

Marcus du Sautoy, mathematician and Simonyi Professor for the Public Understanding of Science at Oxford, doesn’t talk about Iamus in his new book The Creativity Code, but the question he poses about such efforts amounts to this: can we call Iamus a composer? When Iamus’s compositions were played to musically informed listeners, they were unable to distinguish them from music in a similar (modernist) style composed by humans. In a classic 1950 paper, Alan Turing asked: “Can machines think?” If we were unable, in a remote conversation, to distinguish the responses of a person from those of a computer, then we might have to acknowledge that they could. This AI system, then, passes a musical version of the Turing Test.

Du Sautoy’s test is different but no less challenging: can machines be genuinely creative? The interest, just as it was for Turing, lies not so much in finding a definitive answer but in examining what the question itself might mean. No one believes that computer algorithms like Iamus, or any of today’s artificial intelligence systems, really do “think.” They don’t have any consciousness at all. Typically they employ machine learning, in which the algorithm is trained by letting it search for regularities and patterns in a body of “training” data—classical compositions, say—and then apply those rules to generate new “ideas” in the same style.

But this is not so different from how our own cognition works. We learn our native language by experience and inference, not by formal instruction about its rules (although some of that may refine our usage). It’s the same with music, literature and art: we study examples of what others have done, then generalise what we learn and draw on the “rules” in order to produce our own versions.

With machine learning as with human learning, sticking too closely to the training sample tends to generate just more of the same: competent, perhaps, but not very original—or creative. Real advances in the arts arrive when someone ventures outside that sample, disregarding rules or combining them in unexpected ways. This sort of creativity is found equally in the sciences, where machine learning is being increasingly used not just to spot patterns in data that are beyond the capacity of humans to perceive but also to design experiments and even to formulate hypotheses and theories. Du Sautoy is especially eloquent about how it features in mathematics, which he rightfully and persuasively presents as one of the rational human endeavours where creativity is most required.

A common objection is that AI can’t make those conceptual leaps into new territory. But as du Sautoy shows, this need not be true. It is easy to program a bit of rule-breaking randomness into an algorithm, and it is even possible to include an automated means of evaluating the outcomes: searching for quality amongst the quantity. For example, algorithms can be pitted adversarially against one another, one acting as the “thinker” (trained on human judgments, say) that identifies the “good” results produced by the other, thereby leading to gradual, quasi-Darwinian improvement of the outcome. These systems thus imitate the self-critical process of the artist, without the angst. Iamus has such an inner critic, as does a jazz-improvising algorithm called GenBebop, trained by listening to Charlie Parker.

Without such feedback, there’s a danger that AI produces bland and mindless quantity. Du Sautoy describes a program called Mizar, developed by Alphabet (Google’s parent company) and its subsidiary company DeepMind, that generates mathematical theorems by churning through all the possible statements that follow from a given set of axioms. Looking through the proofs Mizar had produced for some of du Sautoy’s favourite theorems, he says “they left me cold.” Therein lies the essence of maths as a human discipline: it is not just about getting a proof, but finding one that pleases the rarefied aesthetic of the discipline, meaning perhaps that it points to other connections, alludes to previous knowledge, while also supplying a payoff “Aha!” moment of insight. All of which is what we appreciate in music too.

What Mizar is producing, says du Sautoy, is akin to Jorge Luis Borges’s Library of Babel, which contains all possible books, made from all possible permutations of characters in all languages. Most are gibberish, but even more importantly, the masterpieces are relegated to just one more product of the random process, of no more consequence to the “creator” than the others. I think du Sautoy is right to suggest that our idea of “creativity is very tied up with mortality”: with finiteness.

But this does not mean that using AI in the creation of works of art, science or maths is doomed and pointless. Du Sautoy argues that it can be exploited to extend human creativity: as a source of basic ideas, raw and unexpected material, that a human might shape and refine. This, indeed, is how some composers are using Iamus. Du Sautoy makes that point in the finest passage in the book, in which he describes how an algorithm produced by DeepMind called AlphaGo—which defeated Lee Sedol—the Korean world champion in the chess-like game Go, has broadened the scope of strategies that human players now use.

And if one day we find ourselves moved by music, poetry or images made by an algorithm, we should not be surprised or embarrassed, nor feel cheated. There are regularities and patterns that stimulate those responses in human-made art, and machine-learning methods are designed to find and exploit them. That there is some predictability to our emotions need not dismay us.

Rather, such responses can deepen our understanding and appreciation of how our own minds work. The problem with many studies of creativity is that they speak of it almost as a substance, a kind of magical essence that can be measured and, if not bought, then acquired by hook or crook. A thing that can be assayed for its purity, like gold. Du Sautoy hints at a better picture: what we perceive as creativity arises from a transaction. We are not passive consumers of the creativity of others. Mozart’s music is all notation and noise until it falls on the ear of the receptive listener. Someone who has never heard western music, or an infant, will struggle to distinguish it from Salieri, or perhaps not perceive it as musical at all.

This interactive aspect is crucial for discussions of “creative” AI. Du Sautoy recounts examples of critics instantly (and anxiously) devaluing their assessments once they learn that a “work of art” was created by an algorithm. It seems to be a fairly universal response, and can’t be dismissed as mere snobbery. Lennox Mackenzie, the LSO’s chairman when the orchestra performed the works of Iamus, confessed that “my normal inclination is to delve into music and find out what it’s all about. But here I don’t think I’d find anything.” We find it harder to take pleasure in a creation devoid of human context or intention—but that of course is contingent knowledge.

Our judgment of creativity depends on a perception of intent. If machines are able to learn to reproduce the surface textures of visual art, music, even poetry and literature, our minds are attuned enough to respond and perhaps to attribute meaning to such works. It is not mere anti-machine prejudice that we should feel our response shift when we discover that nothing more than automated pattern-recognition has created the composition. The common response to computer-generated music or literature—that it is convincing enough in small snatches but can offer no large-scale architecture, no original thesis or involving story—testifies to its lack of a shaping consciousness, and there is no sign yet that computers have anything to offer in its place.

Still, the idea of creativity as a quasi-magical essence that a few privileged—and typically male—individuals dispense to the rest of us is tenacious, and dominates the curiously old-fashioned How To Steal Fire by cultural critic Stephen Bayley and business expert Roger Mavity. This how-to guide comes from a world that believes in “creative types” and holds up Steve Jobs and Elon Musk as the apotheosis of those blessed few. It believes that “genius is raw, not refined,” that all creative geniuses have a touch of madness, that they are unhappy, contrarian, restless souls and that rules only inhibit them.

Scientists generally don’t qualify here; we are told that they don’t like accidents, that they prefer “the illusory sanctuary of numbers” (no one tell du Sautoy) and “the groupthink of peer evaluation.”

Rigour presumably stifles creativity too, so that it’s fine to attribute to famous people like Picasso and Einstein apocryphal quotes that serve your purpose, or to “speculate” that the sexualised women in fashion photographer Helmut Newton’s images are merely “forceful and proud” and “may owe much to the memory of his mother.” (If there’s any truth in that, it is all the more alarming.)

Yet more unforgivable is Bayley’s suggestion that, because of its Confucian heritage, “modern China has yet to innovate in any field.” Not only will any informed artist or scientist know of the creativity now pouring out of China, but the country’s innovation while Confucianism was truly the dominant political philosophy dwarfed that of the west. It’s just the kind of lazy cliché that Bayley excoriates later.

As du Sautoy illustrates, rules can actually be a stimulus to creativity. Not only is there scarcely a more creative sphere than rule-bound mathematics, but artists have drawn inspiration from self-imposed and often highly constraining rules: think of Schoenberg, Mondrian, Ligeti, Seurat, JS Bach.

How To Steal Fire will have none of that. It adopts the “move fast and break things” worldview that appeals to Silicon Valley entrepreneurs and “daring” ad agencies. We’re told that “a creative genius does not submit to ordinary laws: he [sic] writes his own.” The irony is that, while the book rightly dismisses the notion that there is any formula for creativity, at the same time it perpetuates that idea with its dos and don’ts. When it is right (“we need to be wrong more often”; “genius is not the same as wisdom”), it is only recycling platitudes.

This is a carefully curated—and ultra masculine—picture of creativity, where the swaggering Picasso is preferred to the bourgeois Matisse, the bullying, drug-addled Hunter S Thompson to the gently but fiercely thoughtful Ursula Le Guin. “If you observe childlike behaviour in a creative person, forgive them,” writes Bayley. “It’s part of how they are.” In which case, I apologise in advance for any creativity I will be denying the world by an unwillingness to excuse people who act like egotistical jerks.

Bayley and Mavity are also conventional in their gallery of role models: Leonardo, Mozart, Einstein. Yet even if we never agree on what creativity is, we can recognise and celebrate it in other places and at other levels: in home cooking and knitting (there is a history to be told of the creativity that women channelled into the few avenues permitted to most of them), in practical crafts like carpentry and pottery, in games and sport.

I suspect Bayley and Mavity would be appalled at the idea that computer algorithms-—which have never known unhappiness, frustration or bad-boy antics and which even when they break the rules do so according to other rules—might ever be considered creative. If we deem it essential to our own creativity that it involve the purposeful shaping of material in sympathy with other minds, then today’s AI is nowhere close. But if these authors are right that “being able to see a pattern or a possibility invisible to others may be a defining aspect of creativity,” it seems almost inevitable that AI arts will become increasingly difficult to distinguish from the human-made. We might even enjoy them, if we let ourselves.