The week of February 28, 2016

When robots write poetry

By Sasha Chapin

Recently I was in Nepal, talking to a bunch of revolutionaries pursuing a maybe hopeless campaign. “You’re a writer? I could never do that,” one of them asked. I replied that I couldn’t see myself hurling stones at riot police. “But you could do that if you were in my position. Me, I don’t have the writing quality,” he said. Like, the acquired skill of making quality writing? “No, the real quality,” he said, rubbing his fingers together as though grasping an ethereal residue.

This happens frequently. I meet highly accomplished people—oncologists, even—who regard me as some sort of shaman. I see where they’re coming from. As a teenager, it was stupefying to read the creative writers who inspired me. Kurt Vonnegut’s characters, for example—swept away by absurd forces, animated my own inchoate feelings of teenage alienation better than I ever could. It felt like mind-reading.

Writing I’d wincingly but sincerely call “literary”—writing that makes you feel like you’ve found yourself on the page—is an art of unparalleled intimacy. It occurs in private, in shared thought. There’s nothing between writer and reader, just the inner voice being spoken to—and listening.  

When I taught Dylan Thomas to teenagers, one of them wept over “Fern Hill,” saying, “How the fuck did he do that to me?” That’s the feeling I mean—awe and surprise and joy at encountering phrases that both mirror and reform your own feelings of experience, as if you’d encountered your own more eloquent mind.

“A poet is a man who manages, in a lifetime of standing out in thunderstorms, to be struck by lightning five or six times.”

But even Dylan Thomas didn’t manage to do that often. It was hard, hard work that he failed at many times. Many of his poems are straight-up boring—maybe he needed those failures to get to the rare successes. As critic Randall Jarrell wrote, “A poet is a man who manages, in a lifetime of standing out in thunderstorms, to be struck by lightning five or six times.”

Waiting out in the storm for lightning to strike you. That’s the mystery, and the work. But being a full-time writer has dispelled some of the mystery for me. I know my little writing tricks. I can run the “essay about my feelings” algorithm, with varying success. There’s a mystery, yes, but there’s also a recipe.

Then I found the poetry generator. It’s a program designed by biology student Zachary Scholl, capable of producing poems very much like those of one of my all-time favorites, Frank O’Hara.

For reference, here’s some O’Hara, not his best but not his worst:

Oh! kangaroos, sequins, chocolate sodas!
You really are beautiful! Pearls,
harmonicas, jujubes, aspirins! all
the stuff they’ve always talked about
still makes a poem a surprise!

And here are some lines from Scholl’s algorithm:

The enduring dignity of the tiger!
The blood-stained knave is cosmic on your brain.
It magnifies like an alcove inside the pencil.
Sometimes a piece of the clay
dismantles like a jar in my finger.

See what I mean? They have the same swoony quality—the same juxtapositions made palatable by a talky-jokey pace. The algorithm’s lines don’t move me, but they sure are pleasurable.

Scholl wrote his program in 2011. Delighted by what he’d done, he submitted some poems to his school’s literary journal. He also posted them on Reddit. The journal published some, while the redditors were appreciative. His robot’s work passed, at least, for fun juvenilia.

I wondered if Scholl had wanted to fool readers or editors. Nope, he said via email, it wasn’t sinister like that. “My intentions were always based in curiosity,” he said. “I love poetry and I was curious to see if I could make my own poetry by telling my computer how to make it.”

He’d turned Jarrell’s standing in the rain into a process, an algorithm. His favorite poems juxtaposed positive and negative emotions. So he coded a dictionary in which words are scored emotionally: “poison” gets a negative score while “cloud” scores positively, and so on. Those words get slotted into semi-random structures Scholl adapted from poets he admires; the overall score of each sentence should be around zero, with the positive and negative emotions held in counterpoise.

You could complain that a somewhat random collection of vague, floaty intrigue, calibrated for an ambivalent strangeness isn’t the height of poetic accomplishment. But that Scholl could reduce even this kind of poetry to an algorithm suggests that it might not be impossible to build a better poet-bot—one capable of imitating, say, T.S. Eliot.

Eliot, by the way, wouldn’t have been surprised. In his essay “Tradition and the Individual Talent,” he describes writing in a very workmanlike way. You begin by noticing how people regard the history of writing. Then you write stuff that’s reminiscent of that imagined sense of history, spangled with some of the present vernacular. Basically, you regard the poetic tradition as a geometric series, then you simply provide the next iteration.

Eliot didn’t always successfully execute his own formula, but that doesn’t negate his point that poetry can be formulaic. Eliot, though, chose to use the word “tradition” rather than “formula”—the former has a romantic heft, while the latter is dismissive.

But in practice, they’re often synonymous. Discomfort about robots writing poetry might stem from a failure to recognize that. At the same time, part of appreciating writing is an attachment to its human author. You can fall a little in love with the person you imagine based on her sentences. If you knew a poem wasn’t penned by a flesh-and-blood human, but birthed from an algorithm, would you feel the same way?

Perhaps it depends only on the complexity of the algorithm. What happens when an algorithm stops producing pastiches of recognizably human poems and begins to create something truly alien? Conscious machines are coming. Runaway computer intelligence seems inevitable—an AI capable of making itself more intelligent, which then makes itself even smarter, ad infinitum.

“We do not write ant stories for ants with the expectation of being rewarded for our labours with aphid-milked honeydew.”

But the consequences of that won’t be robot Sylvia Plaths. Something much stranger will occur: We’ll have intelligent beings writing for their own motivations rather than ours. It’s a point Peter Swirski makes in From Literature to Biterature: Lem, Turing, Darwin, and Explorations in Computer Literature, Philosophy of Mind, and Cultural Evolution. He asks a simple question: Why would a super-intelligent machine want to write poetry for dumb humans? Maybe an AI would enjoy it like we enjoy scratching a cat’s chin. But we don’t spend all of our time scratching cats’ chins—we like doing difficult, impressive things, like stunning our readers by making laborious artwork. Or, as Swirski puts it, “Let us face it: we do not write ant stories for ants with the expectation of being rewarded for our labours with aphid-milked honeydew.”

So as strange as it might be to feel moved by a robot writer—to find the ghost within the machine—perhaps there’s something infinitely stranger waiting over the horizon. The coming artificial beings may love good poetry for the same reason we do: how it can seem to bridge the boundaries between consciousnesses. But they will possess a consciousness we couldn’t possibly understand. And when they write poetry, it will not be for us.

Illustration via Bruno Moraes