The Future of Life Institute, a think tank in Boston, predicts that artificial intelligence will write bestsellers by 2050, and we’re already on our way. In 2017, Cheers Publishing in China produced Sunshine Misses Windows, a volume of 139 poems created by a Microsoft chatbot that mimics human conversation, or tries to. In short, the computer was fed an image — say, a landscape of rocks, trees, and fog — and instructed to riff. It studied the picture using the GoogleNet algorithm and compared keywords with thousands of previous (albeit human-generated) Chinese poems. The result is a new one: “Wings hold rocks and water tightly / In the loneliness / Stroll the empty / The land becomes soft.”
Robert Lowell was a collector: tools, marbles, butterflies, snakes, turtles, books about Napoleon. “I caught over thirty turtles,” the poet wrote, “and put them in a well where they died of insufficient feeding.” None of this, he admitted, led anywhere. But, of course, that’s not true: he had an instinct for the large sample, for amassing evidence in the hope that something would reveal itself. And, as we know, something did. In this way, Lowell was an algorithm.
It’s been more than forty years since Lowell published his last collection, Day by Day, and today we are “at the end of art, human art” — at least according to the Montreal musician Grimes. Her song “We Appreciate Power” hints at a future for art that will belong to machine learning. AI will control us, she seems to be saying, so it’s best to make friends with it. Grimes, whose real name is Claire Boucher, also says that live music will soon be “obsolete” (anyone who’s paid to see a show involving not much more than a DJ and his laptop might agree). Grimes’s prediction may be tongue-in-cheek experimental dystopianism — a pop star’s marketing gimmick — but others are taking the bait. The American singer Zola Jesus calls Grimes “the voice of silicon fascist privilege,” presumably a dig at Grimes’s billionaire boyfriend, Elon Musk.
But if we want to play along, the idea of the end of human art is a chance to imagine what comes next. What will tomorrow’s bestseller look like? Its bot author might study the history of literature and settle on the story of a vampire whale, set in the American South during Sherman’s March to the Sea, written from the point of view of an autistic child. But that’s fantasy. To ground the project in reality, consider Musk’s thought lab OpenAI, which announced GPT-2 in early 2019. This system used Google Translate to contemplate gigabytes of written material, in order to predict an appropriate continuation of a piece of prose. When GPT-2 was fed “It is a truth universally acknowledged,” the opening to Jane Austen’s Pride and Prejudice, it picked things up from there: “It is a truth universally acknowledged that when a nation is in a condition of civilization, that it is in a great measure the business of its leaders to encourage the habits of virtue, and of industry, and of good order among its people.” What we see, in the end, is a computer struggling with grammar (the Guardian pointed out that redundant second “that”) and (how else to describe it?) with the art part.
Until late last year, OpenAI held off from releasing a full version of GPT-2, due to what Musk called “concerns for malicious applications of the technology,” which may be a self-fulfilling prophecy. Observe how GPT-2 used Austen as a springboard for some wet nonsense about good order and leadership, where a single man in possession of a good fortune must be in want of moral strength, conformity, and trains running on time. Where, ultimately, was it getting its ideas? Is it possible the machine was just telling Musk what he wants to hear, like a writer sucking up to his editor? It wouldn’t be the first time, but that, too, is not art.
The problem may be a brand of AI that looks, simply, to the past as a predictor of what comes next. The musician Holly Herndon, who collaborates with machines, sees the problem running deep in her industry. Warner Music, for example, has Endel, an app that composes mood music. It is marketed as a personalized “sound environment to help you to relax and sleep.” But it also carries an economic bonus. “They’re like: how do I get the system to compose a Hans Zimmer score for me,” Herndon told the Guardian last May, “so that I don’t have to pay an artist?”
So far, AI in the arts seems suited best for those who want to supplement shareholder value — if not actual cultural value. Imagine the computer trying to write like Franz Kafka or Robert Walser. Such artwork is impossible to classify, quantify, or explain, except by pointing to other works or, even better, by reading more Kafka and Walser. One way of looking at how the two writers did what they did is to imagine how much they learned and then unlearned in order to create something fresh. Now, imagine telling Musk’s computer to unlearn what it knows to achieve its goal. Or telling it that poetry, as John Ashbery understood it, is pure language. Pure what? You can practically see smoke coming out of its silicon works, as it tries to learn everything it can about unlearning.
The future of so-called smart machinery in the arts may be, as Herndon has found, one of collaboration. Better yet, it may be doing what we know it does well already: making coffee, just the right strength, for living, breathing poets and composers.